noq_proto/connection/

mod.rs

use std::{
    cmp,
    collections::{BTreeMap, VecDeque, btree_map},
    convert::TryFrom,
    fmt, io, mem,
    net::SocketAddr,
    num::{NonZeroU32, NonZeroUsize},
    sync::Arc,
};

use bytes::{Bytes, BytesMut};
use frame::StreamMetaVec;

use rand::{RngExt, SeedableRng, rngs::StdRng};
use rustc_hash::FxHashMap;
use thiserror::Error;
use tracing::{debug, error, trace, trace_span, warn};

use crate::{
    Dir, Duration, EndpointConfig, FourTuple, Frame, INITIAL_MTU, Instant, MAX_CID_SIZE,
    MAX_STREAM_COUNT, MIN_INITIAL_SIZE, Side, StreamId, TIMER_GRANULARITY, TokenStore, Transmit,
    TransportError, TransportErrorCode, VarInt,
    cid_generator::ConnectionIdGenerator,
    cid_queue::CidQueue,
    config::{ServerConfig, TransportConfig},
    congestion::Controller,
    connection::{
        paths::PathRetransmits,
        qlog::{QlogRecvPacket, QlogSink},
        spaces::LostPacket,
        stats::PathStatsMap,
        timer::{ConnTimer, PathTimer},
    },
    crypto::{self, Keys},
    frame::{
        self, Close, DataBlocked, Datagram, FrameStruct, NewToken, ObservedAddr, StreamDataBlocked,
        StreamsBlocked,
    },
    n0_nat_traversal,
    packet::{
        FixedLengthConnectionIdParser, Header, InitialHeader, InitialPacket, LongType, Packet,
        PacketNumber, PartialDecode, SpaceId,
    },
    range_set::ArrayRangeSet,
    shared::{
        ConnectionEvent, ConnectionEventInner, ConnectionId, DatagramConnectionEvent, EcnCodepoint,
        EndpointEvent, EndpointEventInner,
    },
    token::{ResetToken, Token, TokenPayload},
    transport_parameters::TransportParameters,
};

mod ack_frequency;
use ack_frequency::AckFrequencyState;

mod assembler;
pub use assembler::Chunk;

mod cid_state;
use cid_state::CidState;

mod datagrams;
use datagrams::DatagramState;
pub use datagrams::{Datagrams, SendDatagramError};

mod mtud;
mod pacing;

mod packet_builder;
use packet_builder::{PacketBuilder, PadDatagram};

mod packet_crypto;
use packet_crypto::CryptoState;
pub(crate) use packet_crypto::EncryptionLevel;

mod paths;
pub use paths::{
    ClosedPath, PathAbandonReason, PathEvent, PathId, PathStatus, RttEstimator, SetPathStatusError,
};
use paths::{PathData, PathState};

pub(crate) mod qlog;
pub(crate) mod send_buffer;

pub(crate) mod spaces;
#[cfg(fuzzing)]
pub use spaces::Retransmits;
#[cfg(not(fuzzing))]
use spaces::Retransmits;
pub(crate) use spaces::SpaceKind;
use spaces::{OpenStatus, PacketSpace, SendableFrames, SentPacket, ThinRetransmits};

mod stats;
pub use stats::{ConnectionStats, FrameStats, PathStats, UdpStats};

mod streams;
#[cfg(fuzzing)]
pub use streams::StreamsState;
#[cfg(not(fuzzing))]
use streams::StreamsState;
pub use streams::{
    Chunks, ClosedStream, FinishError, ReadError, ReadableError, RecvStream, SendStream,
    ShouldTransmit, StreamEvent, Streams, WriteError,
};

mod timer;
use timer::{Timer, TimerTable};

mod transmit_buf;
use transmit_buf::TransmitBuf;

mod state;

#[cfg(not(fuzzing))]
use state::State;
#[cfg(fuzzing)]
pub use state::State;
use state::StateType;

/// Protocol state and logic for a single QUIC connection
///
/// Objects of this type receive [`ConnectionEvent`]s and emit [`EndpointEvent`]s and application
/// [`Event`]s to make progress. To handle timeouts, a `Connection` returns timer updates and
/// expects timeouts through various methods. A number of simple getter methods are exposed
/// to allow callers to inspect some of the connection state.
///
/// `Connection` has roughly 4 types of methods:
///
/// - A. Simple getters, taking `&self`
/// - B. Handlers for incoming events from the network or system, named `handle_*`.
/// - C. State machine mutators, for incoming commands from the application. For convenience we
///   refer to this as "performing I/O" below, however as per the design of this library none of the
///   functions actually perform system-level I/O. For example, [`read`](RecvStream::read) and
///   [`write`](SendStream::write), but also things like [`reset`](SendStream::reset).
/// - D. Polling functions for outgoing events or actions for the caller to
///   take, named `poll_*`.
///
/// The simplest way to use this API correctly is to call (B) and (C) whenever
/// appropriate, then after each of those calls, as soon as feasible call all
/// polling methods (D) and deal with their outputs appropriately, e.g. by
/// passing it to the application or by making a system-level I/O call. You
/// should call the polling functions in this order:
///
/// 1. [`poll_transmit`](Self::poll_transmit)
/// 2. [`poll_timeout`](Self::poll_timeout)
/// 3. [`poll_endpoint_events`](Self::poll_endpoint_events)
/// 4. [`poll`](Self::poll)
///
/// Currently the only actual dependency is from (2) to (1), however additional
/// dependencies may be added in future, so the above order is recommended.
///
/// (A) may be called whenever desired.
///
/// Care should be made to ensure that the input events represent monotonically
/// increasing time. Specifically, calling [`handle_timeout`](Self::handle_timeout)
/// with events of the same [`Instant`] may be interleaved in any order with a
/// call to [`handle_event`](Self::handle_event) at that same instant; however
/// events or timeouts with different instants must not be interleaved.
pub struct Connection {
    endpoint_config: Arc<EndpointConfig>,
    config: Arc<TransportConfig>,
    rng: StdRng,
    /// Consolidated cryptographic state
    crypto_state: CryptoState,
    /// The CID we initially chose, for use during the handshake
    handshake_cid: ConnectionId,
    /// The CID the peer initially chose, for use during the handshake
    remote_handshake_cid: ConnectionId,
    /// The [`PathData`] for each path
    ///
    /// This needs to be ordered because [`Connection::poll_transmit`] needs to
    /// deterministically select the next PathId to send on.
    // TODO(flub): well does it really? But deterministic is nice for now.
    paths: BTreeMap<PathId, PathState>,
    /// Counter to uniquely identify every [`PathData`] created in this connection.
    ///
    /// Each [`PathData`] gets a [`PathData::generation`] that is unique among all
    /// [`PathData`]s created in the lifetime of this connection. This helps identify the
    /// correct path when RFC9000-style migrations happen, even when they are
    /// aborted.
    ///
    /// Multipath does not change this, each path can also undergo RFC9000-style
    /// migrations. So a single multipath path ID could see several [`PathData`]s each with
    /// their unique [`PathData::generation].
    path_generation_counter: u64,
    /// Whether MTU detection is supported in this environment
    allow_mtud: bool,
    state: State,
    side: ConnectionSide,
    /// Transport parameters set by the peer
    peer_params: TransportParameters,
    /// Source ConnectionId of the first packet received from the peer
    original_remote_cid: ConnectionId,
    /// Destination ConnectionId sent by the client on the first Initial
    initial_dst_cid: ConnectionId,
    /// The value that the server included in the Source Connection ID field of a Retry packet, if
    /// one was received
    retry_src_cid: Option<ConnectionId>,
    /// Events returned by [`Connection::poll`]
    events: VecDeque<Event>,
    endpoint_events: VecDeque<EndpointEventInner>,
    /// Whether the spin bit is in use for this connection
    spin_enabled: bool,
    /// Outgoing spin bit state
    spin: bool,
    /// Packet number spaces: initial, handshake, 1-RTT
    spaces: [PacketSpace; 3],
    /// Highest usable packet space.
    highest_space: SpaceKind,
    /// Negotiated idle timeout
    idle_timeout: Option<Duration>,
    timers: TimerTable,
    /// Number of packets received which could not be authenticated
    authentication_failures: u64,

    //
    // Queued non-retransmittable 1-RTT data
    //
    /// If the CONNECTION_CLOSE frame needs to be sent
    connection_close_pending: bool,

    //
    // ACK frequency
    //
    ack_frequency: AckFrequencyState,

    //
    // Congestion Control
    //
    /// Whether the most recently received packet had an ECN codepoint set
    receiving_ecn: bool,
    /// Number of packets authenticated
    total_authed_packets: u64,

    //
    // ObservedAddr
    //
    /// Sequence number for the next observed address frame sent to the peer.
    next_observed_addr_seq_no: VarInt,

    streams: StreamsState,
    /// Active and surplus CIDs issued by the remote, for future use on new paths.
    ///
    /// These are given out before multiple paths exist, also for paths that will never
    /// exist.  So if multipath is supported the number of paths here will be higher than
    /// the actual number of paths in use.
    remote_cids: FxHashMap<PathId, CidQueue>,
    /// Attributes of CIDs generated by local endpoint
    ///
    /// Any path that is allowed to be opened is present in this map, as well as the already
    /// opened paths. However since CIDs are issued async by the endpoint driver via
    /// connection events it can not be used to know if CIDs have been issued for a path or
    /// not. See [`Connection::max_path_id_with_cids`] for this.
    local_cid_state: FxHashMap<PathId, CidState>,
    /// State of the unreliable datagram extension
    datagrams: DatagramState,
    /// Path level statistics.
    path_stats: PathStatsMap,
    /// Accumulated stats of all discarded paths.
    ///
    /// The connection-level stats returned by [`Self::stats`] are the sum of the stats of
    /// all the paths. However once a path is discarded it gets added to this field instead
    /// so we do not have to keep an ever growing number of paths stats in memory.
    partial_stats: ConnectionStats,
    /// QUIC version used for the connection.
    version: u32,

    //
    // Multipath
    //
    /// Maximum number of concurrent paths
    ///
    /// Initially set from the [`TransportConfig::max_concurrent_multipath_paths`]. Even
    /// when multipath is disabled this will be set to 1, it is not used in that case
    /// though.
    max_concurrent_paths: NonZeroU32,
    /// Local maximum [`PathId`] to be used
    ///
    /// This is initially set to [`TransportConfig::get_initial_max_path_id`] when multipath
    /// is negotiated, or to [`PathId::ZERO`] otherwise. This is essentially the value of
    /// the highest MAX_PATH_ID frame sent.
    ///
    /// Any path with an ID equal or below this [`PathId`] is either:
    ///
    /// - Abandoned, if it is also in [`Connection::abandoned_paths`].
    /// - Open, in this case it is present in [`Connection::paths`]
    /// - Not yet opened, if it is in neither of these two places.
    ///
    /// Note that for not-yet-open there may or may not be any CIDs issued. See
    /// [`Connection::max_path_id_with_cids`].
    local_max_path_id: PathId,
    /// Remote's maximum [`PathId`] to be used
    ///
    /// This is initially set to the peer's [`TransportParameters::initial_max_path_id`] when
    /// multipath is negotiated, or to [`PathId::ZERO`] otherwise. A peer may increase this limit
    /// by sending [`Frame::MaxPathId`] frames.
    remote_max_path_id: PathId,
    /// The greatest [`PathId`] we have issued CIDs for
    ///
    /// CIDs are only issued for `min(local_max_path_id, remote_max_path_id)`. It is not
    /// possible to use [`Connection::local_cid_state`] to know if CIDs have been issued
    /// since they are issued asynchronously by the endpoint driver.
    max_path_id_with_cids: PathId,
    /// The paths already abandoned
    ///
    /// They may still have some state left in [`Connection::paths`] or
    /// [`Connection::local_cid_state`] since some of this has to be kept around for some
    /// time after a path is abandoned.
    abandoned_paths: AbandonedPaths,

    /// State for n0's (<https://n0.computer>) nat traversal protocol.
    n0_nat_traversal: n0_nat_traversal::State,
    qlog: QlogSink,
}

impl Connection {
    pub(crate) fn new(
        endpoint_config: Arc<EndpointConfig>,
        config: Arc<TransportConfig>,
        init_cid: ConnectionId,
        local_cid: ConnectionId,
        remote_cid: ConnectionId,
        network_path: FourTuple,
        crypto: Box<dyn crypto::Session>,
        cid_gen: &dyn ConnectionIdGenerator,
        now: Instant,
        version: u32,
        allow_mtud: bool,
        rng_seed: [u8; 32],
        side_args: SideArgs,
        qlog: QlogSink,
    ) -> Self {
        let pref_addr_cid = side_args.pref_addr_cid();
        let path_validated = side_args.path_validated();
        let connection_side = ConnectionSide::from(side_args);
        let side = connection_side.side();
        let mut rng = StdRng::from_seed(rng_seed);
        let mut initial_space = PacketSpace::new(now, SpaceId::Initial, &mut rng);
        let mut handshake_space = PacketSpace::new(now, SpaceId::Handshake, &mut rng);
        #[cfg(test)]
        let mut data_space = match config.deterministic_packet_numbers {
            true => PacketSpace::new_deterministic(now, SpaceId::Data),
            false => PacketSpace::new(now, SpaceId::Data, &mut rng),
        };
        #[cfg(not(test))]
        let mut data_space = PacketSpace::new(now, SpaceId::Data, &mut rng);

        // The spaces for PathId::ZERO do not need the PathEvent::Established event.
        initial_space.for_path(PathId::ZERO).open_status = OpenStatus::Informed;
        handshake_space.for_path(PathId::ZERO).open_status = OpenStatus::Informed;
        data_space.for_path(PathId::ZERO).open_status = OpenStatus::Informed;

        let state = State::handshake(state::Handshake {
            remote_cid_set: side.is_server(),
            expected_token: Bytes::new(),
            client_hello: None,
            allow_server_migration: side.is_client() && config.server_handshake_migration,
        });
        let local_cid_state = FxHashMap::from_iter([(
            PathId::ZERO,
            CidState::new(
                cid_gen.cid_len(),
                cid_gen.cid_lifetime(),
                now,
                if pref_addr_cid.is_some() { 2 } else { 1 },
            ),
        )]);

        let mut this = Self {
            endpoint_config,
            crypto_state: CryptoState::new(crypto, init_cid, side, &mut rng),
            handshake_cid: local_cid,
            remote_handshake_cid: remote_cid,
            local_cid_state,
            paths: BTreeMap::from_iter([(
                PathId::ZERO,
                PathState {
                    data: PathData::new(network_path, allow_mtud, None, 0, now, &config),
                    prev: None,
                },
            )]),
            path_generation_counter: 0,
            allow_mtud,
            state,
            side: connection_side,
            peer_params: TransportParameters::default(),
            original_remote_cid: remote_cid,
            initial_dst_cid: init_cid,
            retry_src_cid: None,
            events: VecDeque::new(),
            endpoint_events: VecDeque::new(),
            spin_enabled: config.allow_spin && rng.random_ratio(7, 8),
            spin: false,
            spaces: [initial_space, handshake_space, data_space],
            highest_space: SpaceKind::Initial,
            idle_timeout: match config.max_idle_timeout {
                None | Some(VarInt(0)) => None,
                Some(dur) => Some(Duration::from_millis(dur.0)),
            },
            timers: TimerTable::default(),
            authentication_failures: 0,
            connection_close_pending: false,

            ack_frequency: AckFrequencyState::new(get_max_ack_delay(
                &TransportParameters::default(),
            )),

            receiving_ecn: false,
            total_authed_packets: 0,

            next_observed_addr_seq_no: 0u32.into(),

            streams: StreamsState::new(
                side,
                config.max_concurrent_uni_streams,
                config.max_concurrent_bidi_streams,
                config.send_window,
                config.receive_window,
                config.stream_receive_window,
            ),
            datagrams: DatagramState::default(),
            config,
            remote_cids: FxHashMap::from_iter([(PathId::ZERO, CidQueue::new(remote_cid))]),
            rng,
            path_stats: Default::default(),
            partial_stats: ConnectionStats::default(),
            version,

            // peer params are not yet known, so multipath is not enabled
            max_concurrent_paths: NonZeroU32::MIN,
            local_max_path_id: PathId::ZERO,
            remote_max_path_id: PathId::ZERO,
            max_path_id_with_cids: PathId::ZERO,
            abandoned_paths: Default::default(),

            n0_nat_traversal: Default::default(),
            qlog,
        };
        if path_validated {
            this.on_path_validated(PathId::ZERO);
        }
        if side.is_client() {
            // Kick off the connection
            this.write_crypto();
            this.init_0rtt(now);
        }
        this.qlog
            .emit_tuple_assigned(PathId::ZERO, network_path, now);
        this
    }

    /// Returns the next time at which `handle_timeout` should be called
    ///
    /// The value returned may change after:
    /// - the application performed some I/O on the connection
    /// - a call was made to `handle_event`
    /// - a call to `poll_transmit` returned `Some`
    /// - a call was made to `handle_timeout`
    #[must_use]
    pub fn poll_timeout(&self) -> Option<Instant> {
        self.timers.peek()
    }

    /// Returns application-facing events
    ///
    /// Connections should be polled for events after:
    /// - a call was made to `handle_event`
    /// - a call was made to `handle_timeout`
    #[must_use]
    pub fn poll(&mut self) -> Option<Event> {
        if let Some(x) = self.events.pop_front() {
            return Some(x);
        }

        if let Some(event) = self.streams.poll() {
            return Some(Event::Stream(event));
        }

        if let Some(reason) = self.state.take_error() {
            return Some(Event::ConnectionLost { reason });
        }

        None
    }

    /// Return endpoint-facing events
    #[must_use]
    pub fn poll_endpoint_events(&mut self) -> Option<EndpointEvent> {
        self.endpoint_events.pop_front().map(EndpointEvent)
    }

    /// Provide control over streams
    #[must_use]
    pub fn streams(&mut self) -> Streams<'_> {
        Streams {
            state: &mut self.streams,
            conn_state: &self.state,
        }
    }

    /// Provide control over streams
    #[must_use]
    pub fn recv_stream(&mut self, id: StreamId) -> RecvStream<'_> {
        assert!(id.dir() == Dir::Bi || id.initiator() != self.side.side());
        RecvStream {
            id,
            state: &mut self.streams,
            pending: &mut self.spaces[SpaceId::Data].pending,
        }
    }

    /// Provide control over streams
    #[must_use]
    pub fn send_stream(&mut self, id: StreamId) -> SendStream<'_> {
        assert!(id.dir() == Dir::Bi || id.initiator() == self.side.side());
        SendStream {
            id,
            state: &mut self.streams,
            pending: &mut self.spaces[SpaceId::Data].pending,
            conn_state: &self.state,
        }
    }

    /// Opens a new path only if no path on the same network path currently exists.
    ///
    /// Returns `(path_id, true)` if the path already existed, or `(path_id, false)`
    /// if was opened.
    ///
    /// If `network_path` has no local IP set, then this will open a new path
    /// if no path exists for this remote address, independent of the existing
    /// path's local IP. If a local IP is set, it will match against the full
    /// four-tuple of existing paths. Not setting the local IP avoids having to
    /// guess which local interface will be used to communicate with the remote,
    /// should it not be known yet. We assume that if we already have a path to
    /// the remote, the OS is likely to use the same interface to talk to said remote.
    ///
    /// See also [`open_path`].
    ///
    /// [`open_path`]: Connection::open_path
    pub fn open_path_ensure(
        &mut self,
        network_path: FourTuple,
        initial_status: PathStatus,
        now: Instant,
    ) -> Result<(PathId, bool), PathError> {
        let existing_open_path = self.paths.iter().find(|(id, path)| {
            network_path.is_probably_same_path(&path.data.network_path)
                && !self.abandoned_paths.contains(id)
        });
        match existing_open_path {
            Some((path_id, _state)) => Ok((*path_id, true)),
            None => Ok((self.open_path(network_path, initial_status, now)?, false)),
        }
    }

    /// Opens a new path.
    ///
    /// Further errors might occur and they will be emitted in [`PathEvent::Abandoned`]
    /// events with this path id.  Once the path is opened and can carry application data it
    /// will be reported using a [`PathEvent::Established`] event.
    pub fn open_path(
        &mut self,
        network_path: FourTuple,
        initial_status: PathStatus,
        now: Instant,
    ) -> Result<PathId, PathError> {
        let Some(max_path_id) = self.max_path_id() else {
            return Err(PathError::MultipathNotNegotiated);
        };
        if self.side().is_server() {
            return Err(PathError::ServerSideNotAllowed);
        }

        let max_abandoned = self.abandoned_paths.max();
        let max_used = self.paths.keys().last().copied();
        let path_id = max_abandoned
            .max(max_used)
            .unwrap_or(PathId::ZERO)
            .saturating_add(1u8);

        if path_id > max_path_id {
            self.spaces[SpaceId::Data].pending.paths_blocked = Some(self.remote_max_path_id);
            return Err(PathError::MaxPathIdReached);
        }
        if !self.remote_cids.contains_key(&path_id) {
            self.spaces[SpaceId::Data]
                .pending
                .path_cids_blocked
                .insert(path_id, VarInt(0));
            return Err(PathError::RemoteCidsExhausted);
        }

        let path = self.create_path(path_id, network_path, now, None);
        path.status.local_update(initial_status);

        Ok(path_id)
    }

    /// Closes a path and sends a PATH_ABANDON frame with the passed error code.
    ///
    /// Returns [`ClosePathError::LastOpenPath`] if this is the last open path.
    /// It does allow closing paths which have not yet been opened, as e.g. is the case
    /// when receiving a PATH_ABANDON from the peer for a path that was never opened locally.
    pub fn close_path(
        &mut self,
        now: Instant,
        path_id: PathId,
        error_code: VarInt,
    ) -> Result<(), ClosePathError> {
        self.close_path_inner(
            now,
            path_id,
            PathAbandonReason::ApplicationClosed { error_code },
        )
    }

    /// Closes a path and sends a PATH_ABANDON frame.
    ///
    /// Other than [`Self::close_path`] this allows to specify the reason for the path being closed.
    /// Internally, this should be used over [`Self::close_path`].
    pub(crate) fn close_path_inner(
        &mut self,
        now: Instant,
        path_id: PathId,
        reason: PathAbandonReason,
    ) -> Result<(), ClosePathError> {
        if self.state.is_drained() {
            return Ok(());
        }

        if !self.is_multipath_negotiated() {
            return Err(ClosePathError::MultipathNotNegotiated);
        }
        if self.abandoned_paths.contains(&path_id)
            || Some(path_id) > self.max_path_id()
            || !self.paths.contains_key(&path_id)
        {
            return Err(ClosePathError::ClosedPath);
        }

        let is_last_path = !self
            .paths
            .keys()
            .any(|id| *id != path_id && !self.abandoned_paths.contains(id));

        if is_last_path && !reason.is_remote() {
            return Err(ClosePathError::LastOpenPath);
        }

        self.abandon_path(now, path_id, reason);

        // When the remote abandons our last path, start a grace timer to allow
        // the application to open a replacement path.
        // https://www.ietf.org/archive/id/draft-ietf-quic-multipath-21.html#section-3.4-8
        if is_last_path {
            // The spec suggests 1 PTO, but we use 3 * PTO to account for
            // packet loss when opening a replacement path. Uses initial RTT
            // since the abandoned path's RTT estimate is no longer valid.
            let rtt = RttEstimator::new(self.config.initial_rtt);
            let pto = rtt.pto_base() + self.ack_frequency.max_ack_delay_for_pto();
            let grace = pto * 3;
            self.timers.set(
                Timer::Conn(ConnTimer::NoAvailablePath),
                now + grace,
                self.qlog.with_time(now),
            );
        }

        Ok(())
    }

    /// Unconditionally abandon a path.
    ///
    /// Only to be called once sure this path should be abandoned, all checks
    /// should have happened before calling this.
    fn abandon_path(&mut self, now: Instant, path_id: PathId, reason: PathAbandonReason) {
        trace!(%path_id, ?reason, "abandoning path");

        let pending_space = &mut self.spaces[SpaceId::Data].pending;
        // Send PATH_ABANDON
        pending_space
            .path_abandon
            .insert(path_id, reason.error_code());

        // Remove pending NEW CIDs for this path
        pending_space.new_cids.retain(|cid| cid.path_id != path_id);
        pending_space.path_status.retain(|&id| id != path_id);

        // Cleanup retransmits across ALL paths (CIDs for path_id may have been transmitted on other paths)
        for space in self.spaces[SpaceId::Data].iter_paths_mut() {
            for sent_packet in space.sent_packets.values_mut() {
                if let Some(retransmits) = sent_packet.retransmits.get_mut() {
                    retransmits.new_cids.retain(|cid| cid.path_id != path_id);
                    retransmits.path_status.retain(|&id| id != path_id);
                }
            }
        }

        // We can't send anything on abandoned paths, so we set
        // tail-loss probes to zero.
        // This likely doesn't do much, as the path won't even be tried for sending
        // in poll_transmit after the path is abandoned.
        self.spaces[SpaceId::Data].for_path(path_id).loss_probes = 0;

        // Note: remote CIDs are NOT removed here. They are removed when the PATH_ABANDON
        // frame is actually written to a packet (in populate_packet). This allows sending
        // PATH_ABANDON on the abandoned path itself when no other path exists (#509).
        debug_assert!(!self.state.is_drained()); // requirement for endpoint_events, checked in `close_path_inner`
        self.endpoint_events
            .push_back(EndpointEventInner::RetireResetToken(path_id));

        self.abandoned_paths.insert(path_id);

        for timer in PathTimer::VALUES {
            // match for completeness
            let keep_timer = match timer {
                // These timers deal with sending and receiving PATH_CHALLENGE and
                // PATH_RESPONSE, but now that the path is abandoned, we no longer care about
                // these frames or their timing
                PathTimer::PathValidationFailed | PathTimer::PathChallengeLost => false,
                // These timers deal with the lifetime of the path. Now that the path is abandoned,
                // these are not relevant.
                PathTimer::PathKeepAlive | PathTimer::PathIdle => false,
                // The path has already been informed that outstanding acks should be sent
                // immediately
                PathTimer::MaxAckDelay => false,
                // This timer should not be set, for completeness it's not kept as it's set when
                // the PATH_ABANDON frame is sent.
                PathTimer::PathDrained => false,
                // Sent packets still need to be identified as lost to trigger timely
                // retransmission.
                PathTimer::LossDetection => true,
                // This path should not be used for sending after the PATH_ABANDON frame is sent.
                // However, any outstanding data that should be sent before PATH_ABANDON, should
                // still respect pacing.
                PathTimer::Pacing => true,
            };

            if !keep_timer {
                let qlog = self.qlog.with_time(now);
                self.timers.stop(Timer::PerPath(path_id, timer), qlog);
            }
        }

        // Set the loss detection timer again, as now it should only be set
        // for time-based loss detection, not tail-loss probes, but currently it
        // could still be set to a tail-loss probe.
        // This will reset it to the next time-based loss time, if applicable.
        self.set_loss_detection_timer(now, path_id);

        // Emit event to the application.
        self.events.push_back(Event::Path(PathEvent::Abandoned {
            id: path_id,
            reason,
        }));
    }

    /// Gets the [`PathData`] for a known [`PathId`].
    ///
    /// Will panic if the path_id does not reference any known path.
    #[track_caller]
    fn path_data(&self, path_id: PathId) -> &PathData {
        if let Some(data) = self.paths.get(&path_id) {
            &data.data
        } else {
            panic!(
                "unknown path: {path_id}, currently known paths: {:?}",
                self.paths.keys().collect::<Vec<_>>()
            );
        }
    }

    /// Gets the [`PathData`] for a known [`PathId`].
    ///
    /// Will panic if the path_id does not reference any known path.
    #[track_caller]
    fn path_data_mut(&mut self, path_id: PathId) -> &mut PathData {
        &mut self.paths.get_mut(&path_id).expect("known path").data
    }

    /// Gets a reference to the [`PathData`] for a [`PathId`]
    fn path(&self, path_id: PathId) -> Option<&PathData> {
        self.paths.get(&path_id).map(|path_state| &path_state.data)
    }

    /// Gets a mutable reference to the [`PathData`] for a [`PathId`]
    fn path_mut(&mut self, path_id: PathId) -> Option<&mut PathData> {
        self.paths
            .get_mut(&path_id)
            .map(|path_state| &mut path_state.data)
    }

    /// Returns all known paths.
    ///
    /// There is no guarantee any of these paths are open or usable.
    pub fn paths(&self) -> Vec<PathId> {
        self.paths.keys().copied().collect()
    }

    /// Gets the local [`PathStatus`] for a known [`PathId`]
    pub fn path_status(&self, path_id: PathId) -> Result<PathStatus, ClosedPath> {
        self.path(path_id)
            .map(PathData::local_status)
            .ok_or(ClosedPath { _private: () })
    }

    /// Returns the path's network path represented as a 4-tuple.
    pub fn network_path(&self, path_id: PathId) -> Result<FourTuple, ClosedPath> {
        self.path(path_id)
            .map(|path| path.network_path)
            .ok_or(ClosedPath { _private: () })
    }

    /// Sets the [`PathStatus`] for a known [`PathId`]
    ///
    /// Returns the previous path status on success.
    pub fn set_path_status(
        &mut self,
        path_id: PathId,
        status: PathStatus,
    ) -> Result<PathStatus, SetPathStatusError> {
        if !self.is_multipath_negotiated() {
            return Err(SetPathStatusError::MultipathNotNegotiated);
        }
        let path = self
            .path_mut(path_id)
            .ok_or(SetPathStatusError::ClosedPath)?;
        let prev = match path.status.local_update(status) {
            Some(prev) => {
                self.spaces[SpaceId::Data]
                    .pending
                    .path_status
                    .insert(path_id);
                prev
            }
            None => path.local_status(),
        };
        Ok(prev)
    }

    /// Returns the remote path status
    // TODO(flub): Probably should also be some kind of path event?  Not even sure if I like
    //    this as an API, but for now it allows me to write a test easily.
    // TODO(flub): Technically this should be a Result<Option<PathSTatus>>?
    pub fn remote_path_status(&self, path_id: PathId) -> Option<PathStatus> {
        self.path(path_id).and_then(|path| path.remote_status())
    }

    /// Sets the max_idle_timeout for a specific path.
    ///
    /// The PathIdle timer is immediately re-armed accounting for already-elapsed
    /// idle time. Setting `None` disables the timeout and stops the timer.
    ///
    /// See [`TransportConfig::default_path_max_idle_timeout`] for details.
    ///
    /// Returns the previous value of the setting.
    pub fn set_path_max_idle_timeout(
        &mut self,
        now: Instant,
        path_id: PathId,
        timeout: Option<Duration>,
    ) -> Result<Option<Duration>, ClosedPath> {
        let path = self
            .paths
            .get_mut(&path_id)
            .ok_or(ClosedPath { _private: () })?;
        let prev = mem::replace(&mut path.data.idle_timeout, timeout);

        // Adjust the PathIdle timer, accounting for already-elapsed idle time.
        if !self.state.is_closed() {
            if let Some(new_timeout) = timeout {
                let timer = Timer::PerPath(path_id, PathTimer::PathIdle);
                let deadline = match (prev, self.timers.get(timer)) {
                    (Some(old_timeout), Some(old_deadline)) => {
                        let last_activity = old_deadline.checked_sub(old_timeout).unwrap_or(now);
                        last_activity + new_timeout
                    }
                    _ => now + new_timeout,
                };
                self.timers.set(timer, deadline, self.qlog.with_time(now));
            } else {
                self.timers.stop(
                    Timer::PerPath(path_id, PathTimer::PathIdle),
                    self.qlog.with_time(now),
                );
            }
        }

        Ok(prev)
    }

    /// Sets the keep_alive_interval for a specific path
    ///
    /// See [`TransportConfig::default_path_keep_alive_interval`] for details.
    ///
    /// Returns the previous value of the setting.
    pub fn set_path_keep_alive_interval(
        &mut self,
        path_id: PathId,
        interval: Option<Duration>,
    ) -> Result<Option<Duration>, ClosedPath> {
        let path = self
            .paths
            .get_mut(&path_id)
            .ok_or(ClosedPath { _private: () })?;
        Ok(mem::replace(&mut path.data.keep_alive, interval))
    }

    /// Find an open, validated path that's on the same network path as the given network path.
    ///
    /// Returns the first path matching, even if there's multiple.
    fn find_validated_path_on_network_path(
        &self,
        network_path: FourTuple,
    ) -> Option<(&PathId, &PathState)> {
        self.paths.iter().find(|(path_id, path_state)| {
            path_state.data.validated
                // Would this use the same network path, if network_path were used to send right now?
                && network_path.is_probably_same_path(&path_state.data.network_path)
                && !self.abandoned_paths.contains(path_id)
        })
        // TODO(@divma): we might want to ensure the path has been recently active to consider the
        // address validated
        // matheus23: Perhaps looking at !self.abandoned_paths.contains(path_id) is enough, given keep-alives?
    }

    /// Creates the [`PathData`] for a new [`PathId`].
    ///
    /// Called for incoming packets as well as when opening a new path locally.
    fn create_path(
        &mut self,
        path_id: PathId,
        network_path: FourTuple,
        now: Instant,
        pn: Option<u64>,
    ) -> &mut PathData {
        let valid_path = self.find_validated_path_on_network_path(network_path);
        let validated = valid_path.is_some();
        let initial_rtt = valid_path.map(|(_, path)| path.data.rtt.conservative());
        let vacant_entry = match self.paths.entry(path_id) {
            btree_map::Entry::Vacant(vacant_entry) => vacant_entry,
            btree_map::Entry::Occupied(occupied_entry) => {
                return &mut occupied_entry.into_mut().data;
            }
        };

        debug!(%validated, %path_id, %network_path, "path added");

        // A new path was added. Cancel any pending NoAvailablePath grace timer.
        self.timers.stop(
            Timer::Conn(ConnTimer::NoAvailablePath),
            self.qlog.with_time(now),
        );
        let peer_max_udp_payload_size =
            u16::try_from(self.peer_params.max_udp_payload_size.into_inner()).unwrap_or(u16::MAX);
        self.path_generation_counter = self.path_generation_counter.wrapping_add(1);
        let mut data = PathData::new(
            network_path,
            self.allow_mtud,
            Some(peer_max_udp_payload_size),
            self.path_generation_counter,
            now,
            &self.config,
        );

        data.validated = validated;
        if let Some(initial_rtt) = initial_rtt {
            data.rtt.reset_initial_rtt(initial_rtt);
        }

        // To open a path locally we need to send a packet on the path. Sending a challenge
        // guarantees this.
        data.pending_challenge = true;
        data.pending.observed_address = self
            .config
            .address_discovery_role
            .should_report(&self.peer_params.address_discovery_role);

        let path = vacant_entry.insert(PathState { data, prev: None });

        let mut pn_space = spaces::PacketNumberSpace::new(now, SpaceId::Data, &mut self.rng);
        if let Some(pn) = pn {
            pn_space.dedup.insert(pn);
        }
        self.spaces[SpaceId::Data]
            .number_spaces
            .insert(path_id, pn_space);
        self.qlog.emit_tuple_assigned(path_id, network_path, now);

        // If the remote opened this path we may not have CIDs for it. For locally opened
        // paths the caller should have already made sure we have CIDs and refused to open
        // it if there were none.
        if !self.remote_cids.contains_key(&path_id) {
            debug!(%path_id, "Remote opened path without issuing CIDs");
            self.spaces[SpaceId::Data]
                .pending
                .path_cids_blocked
                .insert(path_id, VarInt(0));
            // Do not abandon this path right away. CIDs might be in-flight still and arrive
            // soon. It is up to the remote to handle this situation.
        }

        &mut path.data
    }

    /// Returns packets to transmit
    ///
    /// Connections should be polled for transmit after:
    /// - the application performed some I/O on the connection
    /// - a call was made to `handle_event`
    /// - a call was made to `handle_timeout`
    ///
    /// `max_datagrams` specifies how many datagrams can be returned inside a
    /// single Transmit using GSO. This must be at least 1.
    #[must_use]
    pub fn poll_transmit(
        &mut self,
        now: Instant,
        max_datagrams: NonZeroUsize,
        buf: &mut Vec<u8>,
    ) -> Option<Transmit> {
        let max_datagrams = match self.config.enable_segmentation_offload {
            false => NonZeroUsize::MIN,
            true => max_datagrams,
        };

        // Each call to poll_transmit can only send datagrams to one destination, because
        // all datagrams in a GSO batch are for the same destination.  Therefore only
        // datagrams for one destination address are produced for each poll_transmit call.

        // Check whether we need to send a close message
        let connection_close_pending = match self.state.as_type() {
            StateType::Drained => {
                for path in self.paths.values_mut() {
                    path.data.app_limited = true;
                }
                return None;
            }
            StateType::Draining | StateType::Closed => {
                // self.connection_close_pending is only reset once the associated packet
                // had been encoded successfully
                if !self.connection_close_pending {
                    for path in self.paths.values_mut() {
                        path.data.app_limited = true;
                    }
                    return None;
                }
                true
            }
            _ => false,
        };

        // Schedule an ACK_FREQUENCY frame if a new one needs to be sent.
        if let Some(config) = &self.config.ack_frequency_config {
            let rtt = self
                .paths
                .values()
                .map(|p| p.data.rtt.get())
                .min()
                .expect("one path exists");
            self.spaces[SpaceId::Data].pending.ack_frequency = self
                .ack_frequency
                .should_send_ack_frequency(rtt, config, &self.peer_params)
                && self.highest_space == SpaceKind::Data
                && self.peer_supports_ack_frequency();
        }

        let mut next_path_id = self.paths.first_entry().map(|e| *e.key());
        while let Some(path_id) = next_path_id {
            if !connection_close_pending
                && let Some(transmit) = self.poll_transmit_off_path(now, buf, path_id)
            {
                #[cfg(test)]
                {
                    self.partial_stats.transmits_tx += 1;
                }
                return Some(transmit);
            }

            let info = self.scheduling_info(path_id);
            if let Some(transmit) = self.poll_transmit_on_path(
                now,
                buf,
                path_id,
                max_datagrams,
                &info,
                connection_close_pending,
            ) {
                #[cfg(test)]
                {
                    self.partial_stats.transmits_tx += 1;
                }
                return Some(transmit);
            }

            // Continue checking other paths, tail-loss probes may need to be sent
            // in all spaces.
            debug_assert!(
                buf.is_empty(),
                "nothing to send on path but buffer not empty"
            );

            next_path_id = self.paths.keys().find(|i| **i > path_id).copied();
        }

        // We didn't produce any application data packet
        debug_assert!(
            buf.is_empty(),
            "there was data in the buffer, but it was not sent"
        );

        if self.state.is_established() {
            // Try MTU probing now
            let mut next_path_id = self.paths.first_entry().map(|e| *e.key());
            while let Some(path_id) = next_path_id {
                if let Some(transmit) = self.poll_transmit_mtu_probe(now, buf, path_id) {
                    #[cfg(test)]
                    {
                        self.partial_stats.transmits_tx += 1;
                    }
                    return Some(transmit);
                }
                next_path_id = self.paths.keys().find(|i| **i > path_id).copied();
            }
        }

        None
    }

    /// Computes the packet scheduling information for this path.
    ///
    /// While this information is only returned for a single path, it is important to know
    /// that this information remains static for the entire span of a single
    /// [`Connection::poll_transmit`] call. In other words, the return value is purely
    /// functional and only depends on the [`PathId`] **during a single** `poll_transmit`
    /// call. It can be computed up-front for all paths but we don't do that because it
    /// involves an allocation.
    ///
    /// See the inline comments for how the  packet scheduling works.
    ///
    /// # Panics
    ///
    /// This will panic if called for a path for which we do not have any [`PathData`], like
    /// so many other functions we have. But this is the only one to document this in its
    /// doc comment. Maybe that should change. Eventually we'll refactor things for this
    /// panic to go away.
    fn scheduling_info(&self, path_id: PathId) -> PathSchedulingInfo {
        // Such a space is preferred for SpaceKind::Data frames.
        let have_validated_status_available_space = self.paths.iter().any(|(path_id, path)| {
            self.remote_cids.contains_key(path_id)
                && !self.abandoned_paths.contains(path_id)
                && path.data.validated
                && path.data.local_status() == PathStatus::Available
        });

        // Such a space is able to send SpaceKind::Data frames.
        let have_validated_space = self.paths.iter().any(|(path_id, path)| {
            self.remote_cids.contains_key(path_id)
                && !self.abandoned_paths.contains(path_id)
                && path.data.validated
        });

        let is_handshaking = self.is_handshaking();
        let has_cids = self.remote_cids.contains_key(&path_id);
        let is_abandoned = self.abandoned_paths.contains(&path_id);
        let path_data = self.path_data(path_id);
        let validated = path_data.validated;
        let status = path_data.local_status();

        // This is the core packet scheduling, whether this space ID may send
        // SpaceKind::Data frames.
        let may_send_data = has_cids
            && !is_abandoned
            && if is_handshaking {
                // There is only one path during the handshake. We want to
                // already send 0-RTT and 0.5-RTT (permitting anti-amplification
                // limit) data.
                true
            } else if !validated {
                // TODO(flub): When we have a network change we might end up
                //    having to abandon all paths and re-open new ones to the
                //    same remotes. This leaves us without any validated
                //    path. Perhaps we should have a way to figure out if the
                //    path is to a previously-validated remote address and allow
                //    sending data to such remotes immediately.
                false
            } else {
                match status {
                    PathStatus::Available => {
                        // Best possible space to send data on.
                        true
                    }
                    PathStatus::Backup => {
                        // If there is a status-available path we prefer that.
                        !have_validated_status_available_space
                    }
                }
            };

        // CONNECTION_CLOSE is allowed to be sent on a non-validated
        // path. Particularly during the handshake we want to send it before the
        // path is validated. Later if there is no validated path available we
        // will also accept sending it on an un-validated path.
        let may_send_close = has_cids
            && !is_abandoned
            && if !validated && have_validated_status_available_space {
                // We have a better space to send on.
                false
            } else {
                // No other validated space, this is as good as it gets.
                true
            };

        // PATH_ABANDON is normally sent together with other SpaceKind::Data frames. But if
        // there is no other validated space to send it on, it can be sent on the path to be
        // abandoned itself if that was validated.
        let may_self_abandon = has_cids && validated && !have_validated_space;

        PathSchedulingInfo {
            is_abandoned,
            may_send_data,
            may_send_close,
            may_self_abandon,
        }
    }

    fn build_transmit(&mut self, path_id: PathId, transmit: TransmitBuf<'_>) -> Transmit {
        debug_assert!(
            !transmit.is_empty(),
            "must not be called with an empty transmit buffer"
        );

        let network_path = self.path_data(path_id).network_path;
        trace!(
            segment_size = transmit.segment_size(),
            last_datagram_len = transmit.len() % transmit.segment_size(),
            %network_path,
            "sending {} bytes in {} datagrams",
            transmit.len(),
            transmit.num_datagrams()
        );
        self.path_data_mut(path_id)
            .inc_total_sent(transmit.len() as u64);

        self.path_stats
            .get_mut(path_id)
            .udp_tx
            .on_sent(transmit.num_datagrams() as u64, transmit.len());

        Transmit {
            destination: network_path.remote,
            size: transmit.len(),
            ecn: if self.path_data(path_id).sending_ecn {
                Some(EcnCodepoint::Ect0)
            } else {
                None
            },
            segment_size: match transmit.num_datagrams() {
                1 => None,
                _ => Some(transmit.segment_size()),
            },
            src_ip: network_path.local_ip,
        }
    }

    /// poll_transmit logic for off-path data.
    fn poll_transmit_off_path(
        &mut self,
        now: Instant,
        buf: &mut Vec<u8>,
        path_id: PathId,
    ) -> Option<Transmit> {
        if let Some(challenge) = self.send_prev_path_challenge(now, buf, path_id) {
            return Some(challenge);
        }
        if let Some(response) = self.send_off_path_path_response(now, buf, path_id) {
            return Some(response);
        }
        if let Some(challenge) = self.send_nat_traversal_path_challenge(now, buf, path_id) {
            return Some(challenge);
        }
        None
    }

    /// poll_transmit logic for on-path data.
    ///
    /// This is not quite the same as for a multipath packet space, since [`PathId::ZERO`]
    /// has 3 packet spaces, which this handles.
    ///
    /// See [`Self::poll_transmit_off_path`] for off-path data.
    #[must_use]
    fn poll_transmit_on_path(
        &mut self,
        now: Instant,
        buf: &mut Vec<u8>,
        path_id: PathId,
        max_datagrams: NonZeroUsize,
        scheduling_info: &PathSchedulingInfo,
        connection_close_pending: bool,
    ) -> Option<Transmit> {
        // Check if there is at least one active CID to use for sending
        let Some(remote_cid) = self.remote_cids.get(&path_id).map(CidQueue::active) else {
            if !self.abandoned_paths.contains(&path_id) {
                debug!(%path_id, "no remote CIDs for path");
            }
            return None;
        };

        // Whether the last packet in the datagram must be padded so the datagram takes up
        // an exact size. An earlier space can decide to not fill an entire datagram and
        // require the next space to fill it further. But may need a specific size of the
        // datagram containing the packet. The final packet built in the datagram must pad
        // to this size.
        let mut pad_datagram = PadDatagram::No;

        // The packet number of the last built packet. This is kept kept across spaces.
        // QUIC is supposed to have a single congestion controller for the Initial,
        // Handshake and Data(PathId::ZERO) spaces.
        let mut last_packet_number = None;

        // If we end up not sending anything, we need to know if that was because there was
        // nothing to send or because we were congestion blocked.
        let mut congestion_blocked = false;

        // Set the segment size to this path's MTU for on-path data.
        let pmtu = self.path_data(path_id).current_mtu().into();
        let mut transmit = TransmitBuf::new(buf, max_datagrams, pmtu);

        // Iterate over the available spaces.
        for space_id in SpaceId::iter() {
            // Only PathId::ZERO uses non Data space ids.
            if path_id != PathId::ZERO && space_id != SpaceId::Data {
                continue;
            }
            match self.poll_transmit_path_space(
                now,
                &mut transmit,
                path_id,
                space_id,
                remote_cid,
                scheduling_info,
                connection_close_pending,
                pad_datagram,
            ) {
                PollPathSpaceStatus::NothingToSend {
                    congestion_blocked: cb,
                } => {
                    congestion_blocked |= cb;
                    // Continue checking other spaces, tail-loss probes may need to be sent
                    // in all spaces.
                }
                PollPathSpaceStatus::WrotePacket {
                    last_packet_number: pn,
                    pad_datagram: pad,
                } => {
                    debug_assert!(!transmit.is_empty(), "transmit must contain packets");
                    last_packet_number = Some(pn);
                    pad_datagram = pad;
                    // Always check higher spaces. If the transmit is full or they have
                    // nothing to send they will not write packets. But if they can, they
                    // must always be allowed to add to this transmit because coalescing may
                    // be required.
                    continue;
                }
                PollPathSpaceStatus::Send {
                    last_packet_number: pn,
                } => {
                    debug_assert!(!transmit.is_empty(), "transmit must contain packets");
                    last_packet_number = Some(pn);
                    break;
                }
            }
        }

        if last_packet_number.is_some() || congestion_blocked {
            self.qlog.emit_recovery_metrics(
                path_id,
                &mut self
                    .paths
                    .get_mut(&path_id)
                    .expect("path_id was iterated from self.paths above")
                    .data,
                now,
            );
        }

        self.path_data_mut(path_id).app_limited =
            last_packet_number.is_none() && !congestion_blocked;

        match last_packet_number {
            Some(last_packet_number) => {
                // Note that when sending in multiple spaces the last packet number will be
                // the one from the highest space.
                self.path_data_mut(path_id).congestion.on_sent(
                    now,
                    transmit.len() as u64,
                    last_packet_number,
                );
                Some(self.build_transmit(path_id, transmit))
            }
            None => None,
        }
    }

    /// poll_transmit logic for a QUIC-MULTIPATH packet number space (PathID + SpaceId).
    #[must_use]
    fn poll_transmit_path_space(
        &mut self,
        now: Instant,
        transmit: &mut TransmitBuf<'_>,
        path_id: PathId,
        space_id: SpaceId,
        remote_cid: ConnectionId,
        scheduling_info: &PathSchedulingInfo,
        // If we need to send a CONNECTION_CLOSE frame.
        connection_close_pending: bool,
        // Whether the current datagram needs to be padded to a certain size.
        mut pad_datagram: PadDatagram,
    ) -> PollPathSpaceStatus {
        // Keep track of the last packet number we wrote. If None we did not write any
        // packets.
        let mut last_packet_number = None;

        // Each loop of this may build one packet. It works logically as follows:
        //
        // - Check if something *needs* to be sent in this space and *can* be sent.
        //   - If not, return to the caller who will call us again for the next space.
        // - Start a new datagram.
        //   - Unless coalescing the packet into an existing datagram.
        // - Write the packet header and payload.
        // - Check if coalescing a next packet into the datagram is possible.
        // - If coalescing, finish packet without padding to leave space in the datagram.
        // - If not coalescing, complete the datagram:
        //   - Finish packet with padding.
        //   - Set the transmit segment size if this is the first datagram.
        // - Loop: next iteration will exit the loop if nothing more to send in this
        //   space. The TransmitBuf will contain a started datagram with space if
        //   coalescing, or completely filled datagram if not coalescing.
        loop {
            // Determine if anything can be sent in this packet number space.
            let max_packet_size = if transmit.datagram_remaining_mut() > 0 {
                // A datagram is started already, we are coalescing another packet into it.
                transmit.datagram_remaining_mut()
            } else {
                // A new datagram needs to be started.
                transmit.segment_size()
            };
            let can_send =
                self.space_can_send(space_id, path_id, max_packet_size, connection_close_pending);
            let needs_loss_probe = self.spaces[space_id].for_path(path_id).loss_probes > 0;
            let space_will_send = {
                if scheduling_info.is_abandoned {
                    // If this path is abandoned then we might still have to send
                    // PATH_ABANDON itself on it if there was no better space
                    // available. Otherwise we want to send the PATH_ABANDON as permitted by
                    // may_send_data however.
                    scheduling_info.may_self_abandon
                        && self.spaces[space_id]
                            .pending
                            .path_abandon
                            .contains_key(&path_id)
                } else if can_send.close && scheduling_info.may_send_close {
                    // This is the best path to send a CONNECTION_CLOSE on.
                    true
                } else if needs_loss_probe || can_send.space_specific {
                    // We always send a loss probe or space-specific frames if the path is
                    // not abandoned.
                    true
                } else {
                    // Anything else we only send if we're the best path for SpaceKind::Data
                    // frames.
                    !can_send.is_empty() && scheduling_info.may_send_data
                }
            };

            if !space_will_send {
                // Nothing more to send. Previous iterations of this loop may have built
                // packets already.
                return match last_packet_number {
                    Some(pn) => PollPathSpaceStatus::WrotePacket {
                        last_packet_number: pn,
                        pad_datagram,
                    },
                    None => {
                        // Only log for spaces which have crypto.
                        if self.crypto_state.has_keys(space_id.encryption_level())
                            || (space_id == SpaceId::Data
                                && self.crypto_state.has_keys(EncryptionLevel::ZeroRtt))
                        {
                            trace!(?space_id, %path_id, "nothing to send in space");
                        }
                        PollPathSpaceStatus::NothingToSend {
                            congestion_blocked: false,
                        }
                    }
                };
            }

            // We want to send on this space, check congestion control if we can. But only
            // if we will need to start a new datagram. If we are coalescing into an already
            // started datagram we do not need to check congestion control again.
            if transmit.datagram_remaining_mut() == 0 {
                let congestion_blocked =
                    self.path_congestion_check(space_id, path_id, transmit, &can_send, now);
                if congestion_blocked != PathBlocked::No {
                    // Previous iterations of this loop may have built packets already.
                    return match last_packet_number {
                        Some(pn) => PollPathSpaceStatus::WrotePacket {
                            last_packet_number: pn,
                            pad_datagram,
                        },
                        None => {
                            return PollPathSpaceStatus::NothingToSend {
                                congestion_blocked: true,
                            };
                        }
                    };
                }

                // If the datagram is full (or there never was one started), we need to start a
                // new one.
                if transmit.num_datagrams() >= transmit.max_datagrams().get() {
                    // No more datagrams allowed.
                    // Previous iterations of this loop may have built packets already.
                    return match last_packet_number {
                        Some(pn) => PollPathSpaceStatus::WrotePacket {
                            last_packet_number: pn,
                            pad_datagram,
                        },
                        None => {
                            return PollPathSpaceStatus::NothingToSend {
                                congestion_blocked: false,
                            };
                        }
                    };
                }

                if needs_loss_probe {
                    // Ensure we have something to send for a tail-loss probe.
                    let request_immediate_ack =
                        space_id == SpaceId::Data && self.peer_supports_ack_frequency();
                    self.spaces[space_id].queue_tail_loss_probe(
                        path_id,
                        request_immediate_ack,
                        &self.streams,
                    );

                    self.spaces[space_id].for_path(path_id).loss_probes -= 1; // needs_loss_probe ensures loss_probes > 0

                    // Clamp the datagram to at most the minimum MTU to ensure that loss
                    // probes can get through and enable recovery even if the path MTU
                    // has shrank unexpectedly.
                    transmit.start_new_datagram_with_size(cmp::min(
                        usize::from(INITIAL_MTU),
                        transmit.segment_size(),
                    ));
                } else {
                    transmit.start_new_datagram();
                }
                trace!(count = transmit.num_datagrams(), "new datagram started");

                // We started a new datagram, we decide later if it needs padding.
                pad_datagram = PadDatagram::No;
            }

            // If coalescing another packet into the existing datagram, there should
            // still be enough space for a whole packet.
            if transmit.datagram_start_offset() < transmit.len() {
                debug_assert!(transmit.datagram_remaining_mut() >= MIN_PACKET_SPACE);
            }

            //
            // From here on, we've determined that a packet will definitely be sent.
            //

            if self.crypto_state.has_keys(EncryptionLevel::Initial)
                && space_id == SpaceId::Handshake
                && self.side.is_client()
            {
                // A client stops both sending and processing Initial packets when it
                // sends its first Handshake packet.
                self.discard_space(now, SpaceKind::Initial);
            }
            if let Some(ref mut prev) = self.crypto_state.prev_crypto {
                prev.update_unacked = false;
            }

            let Some(mut builder) =
                PacketBuilder::new(now, space_id, path_id, remote_cid, transmit, self)
            else {
                // Confidentiality limit is exceeded and the connection has been killed. We
                // should not send any other packets. This works in a roundabout way: We
                // have started a datagram but not written anything into it. So even if we
                // get called again for another space we will see an already started
                // datagram and try and start another packet here. Then be stopped by the
                // same confidentiality limit.
                return PollPathSpaceStatus::NothingToSend {
                    congestion_blocked: false,
                };
            };
            last_packet_number = Some(builder.packet_number);

            if space_id == SpaceId::Initial
                && (self.side.is_client() || can_send.is_ack_eliciting() || needs_loss_probe)
            {
                // https://www.rfc-editor.org/rfc/rfc9000.html#section-14.1
                pad_datagram |= PadDatagram::ToMinMtu;
            }
            if space_id == SpaceId::Data && self.config.pad_to_mtu {
                pad_datagram |= PadDatagram::ToSegmentSize;
            }

            if scheduling_info.may_send_close && can_send.close {
                trace!("sending CONNECTION_CLOSE");
                // Encode ACKs before the ConnectionClose message, to give the receiver
                // a better approximate on what data has been processed. This is
                // especially important with ack delay, since the peer might not
                // have gotten any other ACK for the data earlier on.
                let is_multipath_negotiated = self.is_multipath_negotiated();
                for path_id in self.spaces[space_id]
                    .number_spaces
                    .iter()
                    .filter(|(_, pns)| !pns.pending_acks.ranges().is_empty())
                    .map(|(&path_id, _)| path_id)
                    .collect::<Vec<_>>()
                {
                    Self::populate_acks(
                        now,
                        self.receiving_ecn,
                        path_id,
                        space_id,
                        &mut self.spaces[space_id],
                        is_multipath_negotiated,
                        &mut builder,
                        &mut self.path_stats.get_mut(path_id).frame_tx,
                        self.crypto_state.has_keys(space_id.encryption_level()),
                    );
                }

                // Since there only 64 ACK frames there will always be enough space
                // to encode the ConnectionClose frame too. However we still have the
                // check here to prevent crashes if something changes.

                // TODO(flub): This needs fixing for multipath, to ensure we can always
                //    write the CONNECTION_CLOSE even if we have many PATH_ACKs to send:
                //    https://github.com/n0-computer/noq/issues/367.
                debug_assert!(
                    builder.frame_space_remaining() > frame::ConnectionClose::SIZE_BOUND,
                    "ACKs should leave space for ConnectionClose"
                );
                let stats = &mut self.path_stats.get_mut(path_id).frame_tx;
                if frame::ConnectionClose::SIZE_BOUND < builder.frame_space_remaining() {
                    let max_frame_size = builder.frame_space_remaining();
                    let close: Close = match self.state.as_type() {
                        StateType::Closed => {
                            let reason: Close =
                                self.state.as_closed().expect("checked").clone().into();
                            if space_id == SpaceId::Data || reason.is_transport_layer() {
                                reason
                            } else {
                                TransportError::APPLICATION_ERROR("").into()
                            }
                        }
                        StateType::Draining => TransportError::NO_ERROR("").into(),
                        _ => unreachable!(
                            "tried to make a close packet when the connection wasn't closed"
                        ),
                    };
                    builder.write_frame(close.encoder(max_frame_size), stats);
                }
                let last_pn = builder.packet_number;
                builder.finish_and_track(now, self, path_id, pad_datagram);
                if space_id.kind() == self.highest_space {
                    // Don't send another close packet. Even with multipath we only send
                    // CONNECTION_CLOSE on a single path since we expect our paths to work.
                    self.connection_close_pending = false;
                }
                // Send a close frame in every possible space for robustness, per
                // RFC9000 "Immediate Close during the Handshake". Don't bother trying
                // to send anything else.
                // TODO(flub): This breaks during the handshake if we can not coalesce
                //    packets due to space reasons: the next space would either fail a
                //    debug_assert checking for enough packet space or produce an invalid
                //    packet. We need to keep track of per-space pending CONNECTION_CLOSE to
                //    be able to send these across multiple calls to poll_transmit. Then
                //    check for coalescing space here because initial packets need to be in
                //    padded datagrams. And also add space checks for CONNECTION_CLOSE in
                //    space_can_send so it would stop a GSO batch if the datagram is too
                //    small for another CONNECTION_CLOSE packet.
                return PollPathSpaceStatus::WrotePacket {
                    last_packet_number: last_pn,
                    pad_datagram,
                };
            }

            self.populate_packet(now, space_id, path_id, scheduling_info, &mut builder);

            // ACK-only packets should only be sent when explicitly allowed. If we write them due to
            // any other reason, there is a bug which leads to one component announcing write
            // readiness while not writing any data. This degrades performance. The condition is
            // only checked if the full MTU is available and when potentially large fixed-size
            // frames aren't queued, so that lack of space in the datagram isn't the reason for just
            // writing ACKs.
            debug_assert!(
                !(builder.sent_frames().is_ack_only(&self.streams)
                    && !can_send.acks
                    && (can_send.other || can_send.space_specific)
                    && builder.buf.segment_size()
                        == self.path_data(path_id).current_mtu() as usize
                    && self.datagrams.outgoing.is_empty()),
                "SendableFrames was {can_send:?}, but only ACKs have been written"
            );
            if builder.sent_frames().requires_padding {
                pad_datagram |= PadDatagram::ToMinMtu;
            }

            for path_id in builder.sent_frames().largest_acked.keys() {
                self.spaces[space_id]
                    .for_path(*path_id)
                    .pending_acks
                    .acks_sent();
                self.timers.stop(
                    Timer::PerPath(*path_id, PathTimer::MaxAckDelay),
                    self.qlog.with_time(now),
                );
            }

            // Now we need to finish the packet.  Before we do so we need to know if we will
            // be coalescing the next packet into this one, or will be ending the datagram
            // as well.  Because if this is the last packet in the datagram more padding
            // might be needed because of the packet type, or to fill the GSO segment size.

            let max_packet_size = builder
                .buf
                .datagram_remaining_mut()
                .saturating_sub(builder.predict_packet_end());
            // Are we allowed to coalesce AND is there enough space for another *packet* in
            // this datagram AND will we definitely send another packet?
            if builder.can_coalesce
                && path_id == PathId::ZERO
                && let Some(next_space_id) = space_id.next()
                && max_packet_size > MIN_PACKET_SPACE
                && self
                    .space_can_send(space_id, path_id, max_packet_size, connection_close_pending)
                    .is_empty()
                && self.has_pending_packet(next_space_id, max_packet_size, connection_close_pending)
            {
                // We can append/coalesce the next packet into the current
                // datagram. Finish the current packet without adding extra padding.
                trace!("will coalesce with next packet");
                let last_pn = builder.packet_number;
                builder.finish_and_track(now, self, path_id, PadDatagram::No);
                // We need to return - this loop would re-try the same SpaceId, which we don't want.
                // We want to coalesce with the next space_id.
                return PollPathSpaceStatus::WrotePacket {
                    last_packet_number: last_pn,
                    pad_datagram,
                };
            } else {
                // We need a new datagram for the next packet.  Finish the current
                // packet with padding.
                // TODO(flub): if there isn't any more data to be sent, this will still pad
                //    to the segment size and only discover there is nothing to send before
                //    starting the next packet. That is wasting up to 32 bytes.
                if builder.buf.num_datagrams() > 1 && matches!(pad_datagram, PadDatagram::No) {
                    // If too many padding bytes would be required to continue the
                    // GSO batch after this packet, end the GSO batch here. Ensures
                    // that fixed-size frames with heterogeneous sizes
                    // (e.g. application datagrams) won't inadvertently waste large
                    // amounts of bandwidth. The exact threshold is a bit arbitrary
                    // and might benefit from further tuning, though there's no
                    // universally optimal value.
                    const MAX_PADDING: usize = 32;
                    if builder.buf.datagram_remaining_mut()
                        > builder.predict_packet_end() + MAX_PADDING
                    {
                        trace!(
                            "GSO truncated by demand for {} padding bytes",
                            builder.buf.datagram_remaining_mut() - builder.predict_packet_end()
                        );
                        let last_pn = builder.packet_number;
                        builder.finish_and_track(now, self, path_id, PadDatagram::No);
                        return PollPathSpaceStatus::Send {
                            last_packet_number: last_pn,
                        };
                    }

                    // Pad the current datagram to GSO segment size so it can be
                    // included in the GSO batch.
                    builder.finish_and_track(now, self, path_id, PadDatagram::ToSegmentSize);
                } else {
                    builder.finish_and_track(now, self, path_id, pad_datagram);
                }

                // If this is the first datagram we set the segment size to the size of the
                // first datagram.
                if transmit.num_datagrams() == 1 {
                    transmit.clip_segment_size();
                }
            }
        }
    }

    fn poll_transmit_mtu_probe(
        &mut self,
        now: Instant,
        buf: &mut Vec<u8>,
        path_id: PathId,
    ) -> Option<Transmit> {
        let (active_cid, probe_size) = self.get_mtu_probe_data(now, path_id)?;

        // We are definitely sending a DPLPMTUD probe.
        let mut transmit = TransmitBuf::new(buf, NonZeroUsize::MIN, probe_size as usize);
        transmit.start_new_datagram_with_size(probe_size as usize);

        let mut builder =
            PacketBuilder::new(now, SpaceId::Data, path_id, active_cid, &mut transmit, self)?;

        // We implement MTU probes as ping packets padded up to the probe size
        trace!(?probe_size, "writing MTUD probe");
        builder.write_frame(frame::Ping, &mut self.path_stats.get_mut(path_id).frame_tx);

        // If supported by the peer, we want no delays to the probe's ACK
        if self.peer_supports_ack_frequency() {
            builder.write_frame(
                frame::ImmediateAck,
                &mut self.path_stats.get_mut(path_id).frame_tx,
            );
        }

        builder.finish_and_track(now, self, path_id, PadDatagram::ToSize(probe_size));

        self.path_stats.get_mut(path_id).sent_plpmtud_probes += 1;

        Some(self.build_transmit(path_id, transmit))
    }

    /// Returns the CID and probe size if a DPLPMTUD probe is needed.
    ///
    /// We MTU probe all paths for which all of the following is true:
    /// - We have an active destination CID for the path.
    /// - The remote address *and* path are validated.
    /// - The path is not abandoned.
    /// - The MTU Discovery subsystem wants to probe the path.
    fn get_mtu_probe_data(&mut self, now: Instant, path_id: PathId) -> Option<(ConnectionId, u16)> {
        let active_cid = self.remote_cids.get(&path_id).map(CidQueue::active)?;
        let is_eligible = self.path_data(path_id).validated
            && !self.path_data(path_id).is_validating_path()
            && !self.abandoned_paths.contains(&path_id);

        if !is_eligible {
            return None;
        }
        let next_pn = self.spaces[SpaceId::Data]
            .for_path(path_id)
            .peek_tx_number();
        let probe_size = self
            .path_data_mut(path_id)
            .mtud
            .poll_transmit(now, next_pn)?;

        Some((active_cid, probe_size))
    }

    /// Returns true if there is a further packet to send on [`PathId::ZERO`].
    ///
    /// In other words this is predicting whether the next call to
    /// [`Connection::space_can_send`] issued will return some frames to be sent. Including
    /// having to predict which packet number space it will be invoked with. This depends on
    /// how both [`Connection::poll_transmit_on_path`] and
    /// [`Connection::poll_transmit_path_space`] behave.
    ///
    /// This is needed to determine if packet coalescing can happen. Because the last packet
    /// in a datagram may need to be padded and thus we must know if another packet will
    /// follow or not.
    ///
    /// The next packet can be either in the same space, or in one of the following spaces
    /// on the same path. Because a 0-RTT packet can be coalesced with a 1-RTT packet and
    /// both are in the Data(PathId::ZERO) space. Previous spaces are not checked, because
    /// packets are built from Initial to Handshake to Data spaces.
    fn has_pending_packet(
        &mut self,
        current_space_id: SpaceId,
        max_packet_size: usize,
        connection_close_pending: bool,
    ) -> bool {
        let mut space_id = current_space_id;
        loop {
            let can_send = self.space_can_send(
                space_id,
                PathId::ZERO,
                max_packet_size,
                connection_close_pending,
            );
            if !can_send.is_empty() {
                return true;
            }
            match space_id.next() {
                Some(next_space_id) => space_id = next_space_id,
                None => break,
            }
        }
        false
    }

    /// Checks if creating a new datagram would be blocked by congestion control
    fn path_congestion_check(
        &mut self,
        space_id: SpaceId,
        path_id: PathId,
        transmit: &TransmitBuf<'_>,
        can_send: &SendableFrames,
        now: Instant,
    ) -> PathBlocked {
        // Anti-amplification is only based on `total_sent`, which gets updated after
        // the transmit is sent. Therefore we pass the amount of bytes for datagrams
        // that are already created, as well as 1 byte for starting another datagram. If
        // there is any anti-amplification budget left, we always allow a full MTU to be
        // sent (see https://github.com/quinn-rs/quinn/issues/1082).
        if self.side().is_server()
            && self
                .path_data(path_id)
                .anti_amplification_blocked(transmit.len() as u64 + 1)
        {
            trace!(?space_id, %path_id, "blocked by anti-amplification");
            return PathBlocked::AntiAmplification;
        }

        // Congestion control check.
        // Tail loss probes must not be blocked by congestion, or a deadlock could arise.
        let bytes_to_send = transmit.segment_size() as u64;
        let need_loss_probe = self.spaces[space_id].for_path(path_id).loss_probes > 0;

        if can_send.other && !need_loss_probe && !can_send.close {
            let path = self.path_data(path_id);
            if path.in_flight.bytes + bytes_to_send >= path.congestion.window() {
                trace!(
                    ?space_id,
                    %path_id,
                    in_flight=%path.in_flight.bytes,
                    congestion_window=%path.congestion.window(),
                    "blocked by congestion control",
                );
                return PathBlocked::Congestion;
            }
        }

        // Pacing check.
        if let Some(delay) = self.path_data_mut(path_id).pacing_delay(bytes_to_send, now) {
            let resume_time = now + delay;
            self.timers.set(
                Timer::PerPath(path_id, PathTimer::Pacing),
                resume_time,
                self.qlog.with_time(now),
            );
            // Loss probes and CONNECTION_CLOSE should be subject to pacing, even though
            // they are not congestion controlled.
            trace!(?space_id, %path_id, ?delay, "blocked by pacing");
            return PathBlocked::Pacing;
        }

        PathBlocked::No
    }

    /// Send PATH_CHALLENGE for a previous path if necessary
    ///
    /// QUIC-TRANSPORT section 9.3.3
    /// <https://www.rfc-editor.org/rfc/rfc9000.html#name-off-path-packet-forwarding>
    fn send_prev_path_challenge(
        &mut self,
        now: Instant,
        buf: &mut Vec<u8>,
        path_id: PathId,
    ) -> Option<Transmit> {
        let (prev_cid, prev_path) = self.paths.get_mut(&path_id)?.prev.as_mut()?;
        if !prev_path.pending_challenge {
            return None;
        };
        prev_path.pending_challenge = false;
        let token = self.rng.random();
        let network_path = prev_path.network_path;
        prev_path.record_path_challenge_sent(now, token, network_path);

        debug_assert_eq!(
            self.highest_space,
            SpaceKind::Data,
            "PATH_CHALLENGE queued without 1-RTT keys"
        );
        let buf = &mut TransmitBuf::new(buf, NonZeroUsize::MIN, MIN_INITIAL_SIZE.into());
        buf.start_new_datagram();

        // Use the previous CID to avoid linking the new path with the previous path. We
        // don't bother accounting for possible retirement of that prev_cid because this is
        // sent once, immediately after migration, when the CID is known to be valid. Even
        // if a post-migration packet caused the CID to be retired, it's fair to pretend
        // this is sent first.
        let mut builder = PacketBuilder::new(now, SpaceId::Data, path_id, *prev_cid, buf, self)?;
        let challenge = frame::PathChallenge(token);
        let stats = &mut self.path_stats.get_mut(path_id).frame_tx;
        builder.write_frame_with_log_msg(challenge, stats, Some("validating previous path"));

        // An endpoint MUST expand datagrams that contain a PATH_CHALLENGE frame
        // to at least the smallest allowed maximum datagram size of 1200 bytes,
        // unless the anti-amplification limit for the path does not permit
        // sending a datagram of this size
        builder.pad_to(MIN_INITIAL_SIZE);

        builder.finish(self, now);
        self.path_stats
            .get_mut(path_id)
            .udp_tx
            .on_sent(1, buf.len());

        trace!(
            dst = ?network_path.remote,
            src = ?network_path.local_ip,
            len = buf.len(),
            "sending prev_path off-path challenge",
        );
        Some(Transmit {
            destination: network_path.remote,
            size: buf.len(),
            ecn: None,
            segment_size: None,
            src_ip: network_path.local_ip,
        })
    }

    fn send_off_path_path_response(
        &mut self,
        now: Instant,
        buf: &mut Vec<u8>,
        path_id: PathId,
    ) -> Option<Transmit> {
        let network_path = self
            .paths
            .get_mut(&path_id)
            .map(|state| state.data.network_path)?;
        let cid_queue = self.remote_cids.get_mut(&path_id)?;
        let pns = self.spaces[SpaceKind::Data].for_path(path_id);
        let (token, network_path) = pns.pending_path_responses.pop_off_path(network_path)?;

        // TODO: make off-path probes unlinkable.
        let cid = cid_queue.active();

        // PATH_RESPONSE (off-path)
        let frame = frame::PathResponse(token);

        let buf = &mut TransmitBuf::new(buf, NonZeroUsize::MIN, MIN_INITIAL_SIZE.into());
        buf.start_new_datagram();

        let mut builder = PacketBuilder::new(now, SpaceId::Data, path_id, cid, buf, self)?;
        let stats = &mut self.path_stats.get_mut(path_id).frame_tx;
        builder.write_frame_with_log_msg(frame, stats, Some("(off-path)"));

        // PATH_CHALLENGE (off-path)
        //
        // If we are a client doing NAT traversal, always include a PATH_CHALLENGE with any
        // off-path PATH_RESPONSE. No need to schedule any retries for this, if NAT
        // traversal is taking place then this remote already is being probed with
        // retries, this only speeds up a successful traversal.
        if self
            .find_validated_path_on_network_path(network_path)
            .is_none()
            && self.n0_nat_traversal.client_side().is_ok()
        {
            let token = self.rng.random();
            let stats = &mut self.path_stats.get_mut(path_id).frame_tx;
            builder.write_frame(frame::PathChallenge(token), stats);
            let ip_port = (network_path.remote.ip(), network_path.remote.port());
            self.n0_nat_traversal.mark_probe_sent(ip_port, token);
        }

        // Off-path: not tracked in congestion control. The packet is sent to a
        // different destination than path_id's network path.
        builder.pad_to(MIN_INITIAL_SIZE);
        builder.finish(self, now);

        let size = buf.len();
        self.path_stats.get_mut(path_id).udp_tx.on_sent(1, size);

        trace!(
            dst = ?network_path.remote,
            src = ?network_path.local_ip,
            len = buf.len(),
            "sending off-path PATH_RESPONSE",
        );
        Some(Transmit {
            destination: network_path.remote,
            size,
            ecn: None,
            segment_size: None,
            src_ip: network_path.local_ip,
        })
    }

    /// Send a nat traversal challenge (off-path) on this path if possible.
    fn send_nat_traversal_path_challenge(
        &mut self,
        now: Instant,
        buf: &mut Vec<u8>,
        path_id: PathId,
    ) -> Option<Transmit> {
        let remote = self.n0_nat_traversal.next_probe_addr()?;

        if !self.paths.get(&path_id)?.data.validated {
            // Path is not usable for probing
            return None;
        }

        // TODO: Using the active CID here makes the paths linkable. This is a violation of
        //    RFC9000 but something we want to accept in the short term. Eventually we aim
        //    to fix up the supply of CIDs sufficiently so that we can keep paths unlinkable
        //    again.
        let Some(cid) = self
            .remote_cids
            .get(&path_id)
            .map(|cid_queue| cid_queue.active())
        else {
            trace!(%path_id, "Not sending NAT traversal probe for path with no CIDs");
            return None;
        };
        let token = self.rng.random();

        // PATH_CHALLENGE (NAT probe)
        let frame = frame::PathChallenge(token);

        let mut buf = TransmitBuf::new(buf, NonZeroUsize::MIN, MIN_INITIAL_SIZE.into());
        buf.start_new_datagram();

        let mut builder = PacketBuilder::new(now, SpaceId::Data, path_id, cid, &mut buf, self)?;
        let stats = &mut self.path_stats.get_mut(path_id).frame_tx;
        builder.write_frame_with_log_msg(frame, stats, Some("(nat-traversal)"));
        // Off-path: not tracked in congestion control. The packet is sent to a
        // different destination than path_id's network path.
        builder.finish(self, now);

        // Mark as sent after packet build succeeds.
        self.n0_nat_traversal.mark_probe_sent(remote, token);

        let size = buf.len();
        self.path_stats.get_mut(path_id).udp_tx.on_sent(1, size);

        trace!(dst = ?remote, len = buf.len(), "sending off-path NAT probe");
        Some(Transmit {
            destination: remote.into(),
            size,
            ecn: None,
            segment_size: None,
            src_ip: None,
        })
    }

    /// Indicate what types of frames are ready to send for the given packet number space.
    ///
    /// Only for on-path data.
    ///
    /// *packet_size* is the number of bytes available to build the next packet.
    /// *connection_close_pending* indicates whether a CONNECTION_CLOSE frame needs to be
    /// sent.
    fn space_can_send(
        &mut self,
        space_id: SpaceId,
        path_id: PathId,
        packet_size: usize,
        connection_close_pending: bool,
    ) -> SendableFrames {
        let space = &mut self.spaces[space_id];
        let space_has_crypto = self.crypto_state.has_keys(space_id.encryption_level());

        if !space_has_crypto
            && (space_id != SpaceId::Data
                || !self.crypto_state.has_keys(EncryptionLevel::ZeroRtt)
                || self.side.is_server())
        {
            // Nothing to send in this space
            return SendableFrames::empty();
        }

        let mut can_send = space.can_send(path_id, &self.streams);

        // Check for 1RTT space.
        if space_id == SpaceId::Data {
            let pn = space.for_path(path_id).peek_tx_number();
            // Number of bytes available for frames if this is a 1-RTT packet. We're
            // guaranteed to be able to send an individual frame at least this large in the
            // next 1-RTT packet. This could be generalized to support every space, but it's
            // only needed to handle large fixed-size frames, which only exist in 1-RTT
            // (application datagrams).
            let frame_space_1rtt =
                packet_size.saturating_sub(self.predict_1rtt_overhead(pn, path_id));
            can_send |= self.can_send_1rtt(path_id, frame_space_1rtt);
        }

        can_send.close = connection_close_pending && space_has_crypto;

        can_send
    }

    /// Process `ConnectionEvent`s generated by the associated `Endpoint`
    ///
    /// Will execute protocol logic upon receipt of a connection event, in turn preparing signals
    /// (including application `Event`s, `EndpointEvent`s and outgoing datagrams) that should be
    /// extracted through the relevant methods.
    pub fn handle_event(&mut self, event: ConnectionEvent) {
        use ConnectionEventInner::*;
        match event.0 {
            Datagram(DatagramConnectionEvent {
                now,
                network_path,
                path_id,
                ecn,
                first_decode,
                remaining,
            }) => {
                let span = trace_span!("pkt", %path_id);
                let _guard = span.enter();

                if self.early_discard_packet(network_path, path_id) {
                    // A return value of true indicates we should discard this packet.
                    return;
                }

                let was_anti_amplification_blocked = self
                    .path(path_id)
                    .map(|path| path.anti_amplification_blocked(1))
                    // We never tried to send on an non-existing (new) path so have not been
                    // anti-amplification blocked for it previously.
                    .unwrap_or(false);

                let rx = &mut self.path_stats.get_mut(path_id).udp_rx;
                rx.datagrams += 1;
                rx.bytes += first_decode.len() as u64;
                let data_len = first_decode.len();

                self.handle_decode(now, network_path, path_id, ecn, first_decode);
                // The current `path` might have changed inside `handle_decode` since the packet
                // could have triggered a migration. The packet might also belong to an unknown
                // path and have been rejected. Make sure the data received is accounted for the
                // most recent path by accessing `path` after `handle_decode`.
                if let Some(path) = self.path_mut(path_id) {
                    path.inc_total_recvd(data_len as u64);
                }

                if let Some(data) = remaining {
                    self.path_stats.get_mut(path_id).udp_rx.bytes += data.len() as u64;
                    self.handle_coalesced(now, network_path, path_id, ecn, data);
                }

                if let Some(path) = self.paths.get_mut(&path_id) {
                    self.qlog
                        .emit_recovery_metrics(path_id, &mut path.data, now);
                }

                if was_anti_amplification_blocked {
                    // A prior attempt to set the loss detection timer may have failed due to
                    // anti-amplification, so ensure it's set now. Prevents a handshake deadlock if
                    // the server's first flight is lost.
                    self.set_loss_detection_timer(now, path_id);
                }
            }
            NewIdentifiers(ids, now, cid_len, cid_lifetime) => {
                let path_id = ids.first().map(|issued| issued.path_id).unwrap_or_default();
                debug_assert!(ids.iter().all(|issued| issued.path_id == path_id));

                // Path may have been abandoned while this reply was in flight,
                // retire the CIDs instead of queuing them.
                if self.abandoned_paths.contains(&path_id) {
                    if !self.state.is_drained() {
                        for issued in &ids {
                            self.endpoint_events
                                .push_back(EndpointEventInner::RetireConnectionId(
                                    now,
                                    path_id,
                                    issued.sequence,
                                    false,
                                ));
                        }
                    }
                    return;
                }

                let cid_state = self
                    .local_cid_state
                    .entry(path_id)
                    .or_insert_with(|| CidState::new(cid_len, cid_lifetime, now, 0));
                cid_state.new_cids(&ids, now);

                ids.into_iter().rev().for_each(|frame| {
                    self.spaces[SpaceId::Data].pending.new_cids.push(frame);
                });
                // Always update Timer::PushNewCid
                self.reset_cid_retirement(now);
            }
        }
    }

    /// Returns whether a packet can be discarded early.
    ///
    /// Packets sent on the wrong network path can be entirely ignored, saving further
    /// processing.
    ///
    /// Returns true if a packet coming in for this `path_id` over given `network_path`
    /// should be discarded.
    fn early_discard_packet(&mut self, network_path: FourTuple, path_id: PathId) -> bool {
        if self.is_handshaking() && path_id != PathId::ZERO {
            debug!(%network_path, %path_id, "discarding multipath packet during handshake");
            return true;
        }

        if !self.paths.contains_key(&path_id) && self.abandoned_paths.contains(&path_id) {
            trace!(%path_id, "discarding packet for discarded path");
            return true;
        }

        let peer_may_probe = self.peer_may_probe();
        let local_ip_may_migrate = self.local_ip_may_migrate();

        // If this packet could initiate a migration and we're a client or a server that
        // forbids migration, drop the datagram. This could be relaxed to heuristically
        // permit NAT-rebinding-like migration.
        if let Some(known_path) = self.path_mut(path_id) {
            if network_path.remote != known_path.network_path.remote && !peer_may_probe {
                trace!(
                    %path_id,
                    %network_path,
                    %known_path.network_path,
                    "discarding packet from unrecognized peer"
                );
                return true;
            }

            if known_path.network_path.local_ip.is_some()
                && network_path.local_ip.is_some()
                && known_path.network_path.local_ip != network_path.local_ip
                && !local_ip_may_migrate
            {
                trace!(
                    %path_id,
                    %network_path,
                    %known_path.network_path,
                    "discarding packet sent to incorrect interface"
                );
                return true;
            }
        }
        false
    }

    /// Whether the peer may probe new paths.
    ///
    /// RFC9000 §9 and QNT both have probing packets which may arrive from new paths. This
    /// indicates whether these are allowed or not. This is a strict superset from
    /// [`Self::peer_may_migrate`]: every network path that may be migrated to, may also
    /// be probed. But e.g. servers may not migrate, but can be allowed to probe.
    // TODO(flub): In RFC9000 the server is allowed to send off-path probing packets
    //    once the client has been probing such a 4-tuple. These probes are currently
    //    not yet recognised and will end up being discarded because of this.
    //    See https://github.com/n0-computer/noq/issues/607.
    fn peer_may_probe(&self) -> bool {
        match &self.side {
            ConnectionSide::Client { .. } => {
                if let Some(hs) = self.state.as_handshake() {
                    hs.allow_server_migration
                } else {
                    self.n0_nat_traversal.is_negotiated() && self.is_handshake_confirmed()
                }
            }
            ConnectionSide::Server { server_config } => {
                self.is_handshake_confirmed()
                    && (server_config.migration || self.n0_nat_traversal.is_negotiated())
            }
        }
    }

    /// Whether the peer's remote address may migrate.
    ///
    /// In RFC9000 only the client may migrate.
    ///
    /// QUIC relies on stable endpoints during the handshake. So other than the server's
    /// preferred_address transport parameter no side may migrate before the handshake is
    /// completed.
    ///
    /// It is noteworthy that for iroh we allow server migrations during the handshake when
    /// [`state::Handshake::allow_server_migration`] is enabled, but that is handled earlier
    /// in [`Self::handle_packet`] and without probing the current and previous paths.
    fn peer_may_migrate(&self) -> bool {
        match &self.side {
            ConnectionSide::Server { server_config } => {
                server_config.migration && self.is_handshake_confirmed()
            }
            ConnectionSide::Client { .. } => false,
        }
    }

    /// Whether our local IP address is allowed to change with new incoming packets.
    ///
    /// Incoming packets show us the local IP address we received a packet on, which could
    /// be different from what we thought due to e.g. NAT rebinding or moving from mobile
    /// data to WiFi without being notified of the network change.
    ///
    /// This is only allowed to happen after the handshake is confirmed and when we are the
    /// client. Unless QNT is negotiated in which case the server is also allowed to
    /// migrate.
    ///
    /// Be aware that probing packets, which do not exist in Multipath without QNT, are
    /// exempt from this.
    fn local_ip_may_migrate(&self) -> bool {
        (self.side.is_client() || self.n0_nat_traversal.is_negotiated())
            && self.is_handshake_confirmed()
    }
    /// Process timer expirations
    ///
    /// Executes protocol logic, potentially preparing signals (including application `Event`s,
    /// `EndpointEvent`s and outgoing datagrams) that should be extracted through the relevant
    /// methods.
    ///
    /// It is most efficient to call this immediately after the system clock reaches the latest
    /// `Instant` that was output by `poll_timeout`; however spurious extra calls will simply
    /// no-op and therefore are safe.
    pub fn handle_timeout(&mut self, now: Instant) {
        while let Some((timer, _time)) = self.timers.expire_before(now, &self.qlog) {
            let span = match timer {
                Timer::Conn(timer) => trace_span!("timeout", scope = "conn", ?timer),
                Timer::PerPath(path_id, timer) => {
                    trace_span!("timer_fired", scope="path", %path_id, ?timer)
                }
            };
            let _guard = span.enter();
            trace!("timeout");
            match timer {
                Timer::Conn(timer) => match timer {
                    ConnTimer::Close => {
                        self.state.move_to_drained(None, &mut self.endpoint_events);
                    }
                    ConnTimer::Idle => {
                        self.kill(ConnectionError::TimedOut);
                    }
                    ConnTimer::KeepAlive => {
                        self.ping();
                    }
                    ConnTimer::KeyDiscard => {
                        self.crypto_state.discard_temporary_keys();
                    }
                    ConnTimer::PushNewCid => {
                        while let Some((path_id, when)) = self.next_cid_retirement() {
                            if when > now {
                                break;
                            }
                            match self.local_cid_state.get_mut(&path_id) {
                                None => error!(%path_id, "No local CID state for path"),
                                Some(cid_state) => {
                                    // Update `retire_prior_to` field in NEW_CONNECTION_ID frame
                                    let num_new_cid = cid_state.on_cid_timeout().into();
                                    if !self.state.is_closed() {
                                        trace!(
                                            "push a new CID to peer RETIRE_PRIOR_TO field {}",
                                            cid_state.retire_prior_to()
                                        );
                                        self.endpoint_events.push_back(
                                            EndpointEventInner::NeedIdentifiers(
                                                path_id,
                                                now,
                                                num_new_cid,
                                            ),
                                        );
                                    }
                                }
                            }
                        }
                    }
                    ConnTimer::NoAvailablePath => {
                        // Grace period expired: all paths were abandoned and no new path
                        // was opened. Close the connection. There are no paths left to
                        // send CONNECTION_CLOSE on, so this is a silent close.
                        // https://www.ietf.org/archive/id/draft-ietf-quic-multipath-21.html#section-3.4-8
                        if self.state.is_closed() || self.state.is_drained() {
                            // Connection already closing/drained (e.g. application called
                            // close() before the grace timer fired). Nothing to do.
                            error!("no viable path timer fired, but connection already closing");
                        } else {
                            trace!("no viable path grace period expired, closing connection");
                            let err = TransportError::NO_VIABLE_PATH(
                                "last path abandoned, no new path opened",
                            );
                            self.close_common();
                            self.set_close_timer(now);
                            self.connection_close_pending = true;
                            self.state.move_to_closed(err);
                        }
                    }
                    ConnTimer::NatTraversalProbeRetry => {
                        self.n0_nat_traversal.queue_retries(self.is_ipv6());
                        if let Some(delay) =
                            self.n0_nat_traversal.retry_delay(self.config.initial_rtt)
                        {
                            self.timers.set(
                                Timer::Conn(ConnTimer::NatTraversalProbeRetry),
                                now + delay,
                                self.qlog.with_time(now),
                            );
                            trace!("re-queued NAT probes");
                        } else {
                            trace!("no more NAT probes remaining");
                        }
                    }
                },
                Timer::PerPath(path_id, timer) => {
                    match timer {
                        PathTimer::PathIdle => {
                            if let Err(err) =
                                self.close_path_inner(now, path_id, PathAbandonReason::TimedOut)
                            {
                                warn!(?err, "failed closing path");
                            }
                        }

                        PathTimer::PathKeepAlive => {
                            self.ping_path(path_id).ok();
                        }
                        PathTimer::LossDetection => {
                            self.on_loss_detection_timeout(now, path_id);
                            if let Some(path) = self.paths.get_mut(&path_id) {
                                self.qlog
                                    .emit_recovery_metrics(path_id, &mut path.data, now);
                            } else {
                                error!("LossDetection fired for unknown path");
                            }
                        }
                        PathTimer::PathValidationFailed => {
                            let Some(path) = self.paths.get_mut(&path_id) else {
                                continue;
                            };
                            self.timers.stop(
                                Timer::PerPath(path_id, PathTimer::PathChallengeLost),
                                self.qlog.with_time(now),
                            );
                            debug!("path migration validation failed");
                            path.data.reset_on_path_challenges();
                            if let Some((_, prev)) = path.prev.take() {
                                path.data = prev;
                                self.set_loss_detection_timer(now, path_id);
                            }
                        }
                        PathTimer::PathChallengeLost => {
                            let Some(path) = self.paths.get_mut(&path_id) else {
                                continue;
                            };
                            trace!(?path.data.lost_challenge_count, "path challenge deemed lost");
                            path.data.pending_challenge = true;
                            path.data.lost_challenge_count += 1;
                            self.timers.set(
                                Timer::PerPath(path_id, PathTimer::PathChallengeLost),
                                now + path.data.on_path_challenge_pto(),
                                self.qlog.with_time(now),
                            );
                        }
                        PathTimer::Pacing => {}
                        PathTimer::MaxAckDelay => {
                            // This timer is only armed in the Data space
                            self.spaces[SpaceId::Data]
                                .for_path(path_id)
                                .pending_acks
                                .on_max_ack_delay_timeout()
                        }
                        PathTimer::PathDrained => {
                            // The path was abandoned and 3*PTO has expired since.  Clean up all
                            // remaining state and install stateless reset token.
                            self.timers.stop_per_path(path_id, self.qlog.with_time(now));
                            if let Some(local_cid_state) = self.local_cid_state.remove(&path_id) {
                                debug_assert!(!self.state.is_drained()); // requirement for endpoint_events. All timers should be cleared in drained connections.
                                let (min_seq, max_seq) = local_cid_state.active_seq();
                                for seq in min_seq..=max_seq {
                                    self.endpoint_events.push_back(
                                        EndpointEventInner::RetireConnectionId(
                                            now, path_id, seq, false,
                                        ),
                                    );
                                }
                            }
                            self.discard_path(path_id, now);
                        }
                    }
                }
            }
        }
    }

    /// Close a connection immediately
    ///
    /// This does not ensure delivery of outstanding data. It is the application's responsibility to
    /// call this only when all important communications have been completed, e.g. by calling
    /// [`SendStream::finish`] on outstanding streams and waiting for the corresponding
    /// [`StreamEvent::Finished`] event.
    ///
    /// If [`Streams::send_streams`] returns 0, all outstanding stream data has been
    /// delivered. There may still be data from the peer that has not been received.
    ///
    /// [`StreamEvent::Finished`]: crate::StreamEvent::Finished
    pub fn close(&mut self, now: Instant, error_code: VarInt, reason: Bytes) {
        self.close_inner(
            now,
            Close::Application(frame::ApplicationClose { error_code, reason }),
        )
    }

    /// Close the connection immediately, initiated by an API call.
    ///
    /// This will not produce a [`ConnectionLost`] event propagated by the
    /// [`Connection::poll`] call, because the API call already propagated the error to the
    /// user.
    ///
    /// Not to be used when entering immediate close due to an internal state change based
    /// on an event. See [`State::move_to_closed_local`] for details.
    ///
    /// This initiates immediate close from
    /// <https://www.rfc-editor.org/rfc/rfc9000.html#section-10.2>, moving to the closed
    /// state.
    ///
    /// [`ConnectionLost`]: crate::Event::ConnectionLost
    /// [`Connection::poll`]: super::Connection::poll
    fn close_inner(&mut self, now: Instant, reason: Close) {
        let was_closed = self.state.is_closed();
        if !was_closed {
            self.close_common();
            self.set_close_timer(now);
            self.connection_close_pending = true;
            self.state.move_to_closed_local(reason);
        }
    }

    /// Control datagrams
    pub fn datagrams(&mut self) -> Datagrams<'_> {
        Datagrams { conn: self }
    }

    /// Returns connection statistics
    pub fn stats(&mut self) -> ConnectionStats {
        let mut stats = self.partial_stats.clone();

        for path_stats in self.path_stats.iter_stats() {
            // Self::path_stats() computes the path rtt, cwnd and current_mtu on access
            // because they are not simple counters. When computing the connection stats we
            // can skip that effort since those fields are not used in the `impl
            // Add<PathStats> for ConnectionStats`.
            stats += *path_stats;
        }

        stats
    }

    /// Returns path statistics
    pub fn path_stats(&mut self, path_id: PathId) -> Option<PathStats> {
        let path = self.paths.get(&path_id)?;
        let mut stats = self.path_stats.get(path_id).unwrap_or_default();
        stats.rtt = path.data.rtt.get();
        stats.cwnd = path.data.congestion.window();
        stats.current_mtu = path.data.mtud.current_mtu();
        Some(stats)
    }

    /// Ping the remote endpoint
    ///
    /// Causes an ACK-eliciting packet to be transmitted on the connection.
    pub fn ping(&mut self) {
        // TODO(flub): This is very brute-force: it pings *all* the paths.  Instead it would
        //    be nice if we could only send a single packet for this.
        for path_data in self.spaces[self.highest_space].number_spaces.values_mut() {
            path_data.pending_ping = true;
        }
    }

    /// Ping the remote endpoint over a specific path
    ///
    /// Causes an ACK-eliciting packet to be transmitted on the path.
    pub fn ping_path(&mut self, path: PathId) -> Result<(), ClosedPath> {
        let path_data = self.spaces[self.highest_space]
            .number_spaces
            .get_mut(&path)
            .ok_or(ClosedPath { _private: () })?;
        path_data.pending_ping = true;
        Ok(())
    }

    /// Update traffic keys spontaneously
    ///
    /// This can be useful for testing key updates, as they otherwise only happen infrequently.
    pub fn force_key_update(&mut self) {
        if !self.state.is_established() {
            debug!("ignoring forced key update in illegal state");
            return;
        }
        if self.crypto_state.prev_crypto.is_some() {
            // We already just updated, or are currently updating, the keys. Concurrent key updates
            // are illegal.
            debug!("ignoring redundant forced key update");
            return;
        }
        self.crypto_state.update_keys(None, false);
    }

    /// Get a session reference
    pub fn crypto_session(&self) -> &dyn crypto::Session {
        self.crypto_state.session.as_ref()
    }

    /// Whether the connection is in the process of being established
    ///
    /// If this returns `false`, the connection may be either established or closed, signaled by the
    /// emission of a [`Connected`](Event::Connected) or [`ConnectionLost`](Event::ConnectionLost)
    /// event respectively. Note that locally-initiated closes via [`close()`](Self::close) do not
    /// emit a `ConnectionLost` event.
    ///
    /// For an established connection this essentially means the handshake is **completed**,
    /// but not necessarily yet confirmed.
    pub fn is_handshaking(&self) -> bool {
        self.state.is_handshake()
    }

    /// Whether the connection is closed
    ///
    /// Closed connections cannot transport any further data. A connection becomes closed when
    /// either peer application intentionally closes it, or when either transport layer detects an
    /// error such as a time-out or certificate validation failure.
    ///
    /// A [`ConnectionLost`](Event::ConnectionLost) event is emitted with details when the
    /// connection is closed by the peer or due to an error. When the local application closes
    /// the connection via [`close()`](Self::close), no `ConnectionLost` event is emitted;
    /// instead, pending operations fail with [`ConnectionError::LocallyClosed`].
    pub fn is_closed(&self) -> bool {
        self.state.is_closed()
    }

    /// Whether there is no longer any need to keep the connection around
    ///
    /// Closed connections become drained after a brief timeout to absorb any remaining in-flight
    /// packets from the peer. All drained connections have been closed.
    pub fn is_drained(&self) -> bool {
        self.state.is_drained()
    }

    /// For clients, if the peer accepted the 0-RTT data packets
    ///
    /// The value is meaningless until after the handshake completes.
    pub fn accepted_0rtt(&self) -> bool {
        self.crypto_state.accepted_0rtt
    }

    /// Whether 0-RTT is/was possible during the handshake
    pub fn has_0rtt(&self) -> bool {
        self.crypto_state.zero_rtt_enabled
    }

    /// Whether there are any pending retransmits
    pub fn has_pending_retransmits(&self) -> bool {
        !self.spaces[SpaceId::Data].pending.is_empty(&self.streams)
    }

    /// Look up whether we're the client or server of this Connection
    pub fn side(&self) -> Side {
        self.side.side()
    }

    /// Get the address observed by the remote over the given path
    pub fn path_observed_address(&self, path_id: PathId) -> Result<Option<SocketAddr>, ClosedPath> {
        self.path(path_id)
            .map(|path_data| {
                path_data
                    .last_observed_addr_report
                    .as_ref()
                    .map(|observed| observed.socket_addr())
            })
            .ok_or(ClosedPath { _private: () })
    }

    /// Current best estimate of this connection's latency (round-trip-time)
    pub fn rtt(&self, path_id: PathId) -> Option<Duration> {
        self.path(path_id).map(|d| d.rtt.get())
    }

    /// Current state of this connection's congestion controller, for debugging purposes
    pub fn congestion_state(&self, path_id: PathId) -> Option<&dyn Controller> {
        self.path(path_id).map(|d| d.congestion.as_ref())
    }

    /// Modify the number of remotely initiated streams that may be concurrently open
    ///
    /// No streams may be opened by the peer unless fewer than `count` are already open. Large
    /// `count`s increase both minimum and worst-case memory consumption.
    pub fn set_max_concurrent_streams(&mut self, dir: Dir, count: VarInt) {
        self.streams.set_max_concurrent(dir, count);
        // If the limit was reduced, then a flow control update previously deemed insignificant may
        // now be significant.
        let pending = &mut self.spaces[SpaceId::Data].pending;
        self.streams.queue_max_stream_id(pending);
    }

    /// Modify the number of open paths allowed when multipath is enabled
    ///
    /// When reducing the number of concurrent paths this will only affect delaying sending
    /// new MAX_PATH_ID frames until fewer than this number of paths are possible.  To
    /// actively reduce paths they must be closed using [`Connection::close_path`], which
    /// can also be used to close not-yet-opened paths.
    ///
    /// If multipath is not negotiated (see the [`TransportConfig`]) this can not enable
    /// multipath and will fail.
    pub fn set_max_concurrent_paths(
        &mut self,
        now: Instant,
        count: NonZeroU32,
    ) -> Result<(), MultipathNotNegotiated> {
        if !self.is_multipath_negotiated() {
            return Err(MultipathNotNegotiated { _private: () });
        }
        self.max_concurrent_paths = count;

        let in_use_count = self
            .local_max_path_id
            .next()
            .saturating_sub(self.abandoned_paths.len())
            .as_u32();
        let extra_needed = count.get().saturating_sub(in_use_count);
        let new_max_path_id = self.local_max_path_id.saturating_add(extra_needed);

        self.set_max_path_id(now, new_max_path_id);

        Ok(())
    }

    /// If needed, issues a new MAX_PATH_ID frame and new CIDs for any newly allowed paths
    fn set_max_path_id(&mut self, now: Instant, max_path_id: PathId) {
        if max_path_id <= self.local_max_path_id {
            return;
        }

        self.local_max_path_id = max_path_id;
        self.spaces[SpaceId::Data].pending.max_path_id = true;

        self.issue_first_path_cids(now);
    }

    /// Current number of remotely initiated streams that may be concurrently open
    ///
    /// If the target for this limit is reduced using [`set_max_concurrent_streams`](Self::set_max_concurrent_streams),
    /// it will not change immediately, even if fewer streams are open. Instead, it will
    /// decrement by one for each time a remotely initiated stream of matching directionality is closed.
    pub fn max_concurrent_streams(&self, dir: Dir) -> u64 {
        self.streams.max_concurrent(dir)
    }

    /// See [`TransportConfig::send_window()`]
    pub fn set_send_window(&mut self, send_window: u64) {
        self.streams.set_send_window(send_window);
    }

    /// See [`TransportConfig::receive_window()`]
    pub fn set_receive_window(&mut self, receive_window: VarInt) {
        if self.streams.set_receive_window(receive_window) {
            self.spaces[SpaceId::Data].pending.max_data = true;
        }
    }

    /// Whether the Multipath for QUIC extension is enabled.
    ///
    /// Multipath is only enabled after the handshake is completed and if it was enabled by both
    /// peers.
    pub fn is_multipath_negotiated(&self) -> bool {
        !self.is_handshaking()
            && self.config.max_concurrent_multipath_paths.is_some()
            && self.peer_params.initial_max_path_id.is_some()
    }

    fn on_ack_received(
        &mut self,
        now: Instant,
        space: SpaceId,
        ack: frame::Ack,
    ) -> Result<(), TransportError> {
        // All ACKs are referencing path 0
        let path = PathId::ZERO;
        self.inner_on_ack_received(now, space, path, ack)
    }

    fn on_path_ack_received(
        &mut self,
        now: Instant,
        space: SpaceId,
        path_ack: frame::PathAck,
    ) -> Result<(), TransportError> {
        let (ack, path) = path_ack.into_ack();
        self.inner_on_ack_received(now, space, path, ack)
    }

    /// Handles an ACK frame acknowledging packets sent on *path*.
    fn inner_on_ack_received(
        &mut self,
        now: Instant,
        space: SpaceId,
        path: PathId,
        ack: frame::Ack,
    ) -> Result<(), TransportError> {
        if !self.spaces[space].number_spaces.contains_key(&path) {
            if self.abandoned_paths.contains(&path) {
                // See also
                // https://www.ietf.org/archive/id/draft-ietf-quic-multipath-21.html#section-3.4.3-3
                // > When an endpoint finally deletes all state associated with the path [...]
                // > PATH_ACK frames received with an abandoned path ID are silently ignored,
                // > as specified in Section 4.
                trace!("silently ignoring PATH_ACK on discarded path");
                return Ok(());
            } else {
                return Err(TransportError::PROTOCOL_VIOLATION(
                    "received PATH_ACK with path ID never used",
                ));
            }
        }
        if ack.largest >= self.spaces[space].for_path(path).next_packet_number {
            return Err(TransportError::PROTOCOL_VIOLATION("unsent packet acked"));
        }
        // `Some(pn)` if this ACK raised `largest_acked_packet_pn`.
        let new_largest_pn = {
            let space = &mut self.spaces[space].for_path(path);
            if space
                .largest_acked_packet_pn
                .is_none_or(|pn| ack.largest > pn)
            {
                space.largest_acked_packet_pn = Some(ack.largest);
                if let Some(info) = space.sent_packets.get(ack.largest) {
                    // This should always succeed, but a misbehaving peer might ACK a packet we
                    // haven't sent. At worst, that will result in us spuriously reducing the
                    // congestion window.
                    space.largest_acked_packet_send_time = info.time_sent;
                }
                Some(ack.largest)
            } else {
                None
            }
        };

        if self.detect_spurious_loss(&ack, space, path) {
            self.path_stats.get_mut(path).spurious_congestion_events += 1;
            self.path_data_mut(path)
                .congestion
                .on_spurious_congestion_event();
        }

        // Avoid DoS from unreasonably huge ack ranges by filtering out just the new acks.
        let mut newly_acked: ArrayRangeSet = ArrayRangeSet::new();
        for range in ack.iter() {
            self.spaces[space].for_path(path).check_ack(range.clone())?;
            for (pn, _) in self.spaces[space]
                .for_path(path)
                .sent_packets
                .iter_range(range)
            {
                newly_acked.insert_one(pn);
            }
        }

        if newly_acked.is_empty() {
            return Ok(());
        }

        let mut ack_eliciting_acked = false;
        for packet in newly_acked.elts() {
            if let Some(info) = self.spaces[space].for_path(path).take(packet) {
                for (acked_path_id, acked_pn) in info.largest_acked.iter() {
                    // Assume ACKs for all packets below the largest acknowledged in
                    // `packet` have been received. This can cause the peer to spuriously
                    // retransmit if some of our earlier ACKs were lost, but allows for
                    // simpler state tracking. See discussion at
                    // https://www.rfc-editor.org/rfc/rfc9000.html#name-limiting-ranges-by-tracking
                    if let Some(pns) = self.spaces[space].path_space_mut(*acked_path_id) {
                        pns.pending_acks.subtract_below(*acked_pn);
                    }
                }
                ack_eliciting_acked |= info.ack_eliciting;

                // Notify MTU discovery that a packet was acked, because it might be an MTU probe
                let path_data = self.path_data_mut(path);
                let mtu_updated = path_data.mtud.on_acked(space.kind(), packet, info.size);
                if mtu_updated {
                    path_data
                        .congestion
                        .on_mtu_update(path_data.mtud.current_mtu());
                }

                // Notify ack frequency that a packet was acked, because it might contain an ACK_FREQUENCY frame
                self.ack_frequency.on_acked(path, packet);

                self.on_packet_acked(now, path, packet, info);
            }
        }

        let largest_ackd = self.spaces[space].for_path(path).largest_acked_packet_pn;
        let path_data = self.path_data_mut(path);
        let app_limited = path_data.app_limited;
        let in_flight = path_data.in_flight.bytes;

        path_data
            .congestion
            .on_end_acks(now, in_flight, app_limited, largest_ackd);

        if new_largest_pn.is_some() && ack_eliciting_acked {
            let ack_delay = if space != SpaceId::Data {
                Duration::from_micros(0)
            } else {
                cmp::min(
                    self.ack_frequency.peer_max_ack_delay,
                    Duration::from_micros(ack.delay << self.peer_params.ack_delay_exponent.0),
                )
            };
            let rtt = now.saturating_duration_since(
                self.spaces[space]
                    .for_path(path)
                    .largest_acked_packet_send_time,
            );

            let next_pn = self.spaces[space].for_path(path).next_packet_number;
            let path_data = self.path_data_mut(path);
            // TODO(@divma): should be a method of path, should be contained in a single place
            path_data.rtt.update(ack_delay, rtt);
            if path_data.first_packet_after_rtt_sample.is_none() {
                path_data.first_packet_after_rtt_sample = Some((space.kind(), next_pn));
            }
        }

        // Must be called before crypto/pto_count are clobbered
        self.detect_lost_packets(now, space, path, true);

        // If the peer did not complete the handshake address validation the ACK could be
        // spoofed, e.g. in the Initial space. Setting the pto_count back to 0 removes the
        // exponential backoff from the PTO timer and would result in too many tail-loss
        // probes being sent.
        if self.peer_completed_handshake_address_validation() {
            self.path_data_mut(path).pto_count = 0;
        }

        // Explicit congestion notification
        // TODO(@divma): this code is a good example of logic that should be contained in a single
        // place but it's split between the path data and the packet number space data, we should
        // find a way to make this work without two lookups
        if self.path_data(path).sending_ecn {
            if let Some(ecn) = ack.ecn {
                // We only examine ECN counters from ACKs that we are certain we received in transmit
                // order, allowing us to compute an increase in ECN counts to compare against the number
                // of newly acked packets that remains well-defined in the presence of arbitrary packet
                // reordering.
                if let Some(largest_sent_pn) = new_largest_pn {
                    let sent = self.spaces[space]
                        .for_path(path)
                        .largest_acked_packet_send_time;
                    self.process_ecn(
                        now,
                        space,
                        path,
                        newly_acked.range_count() as u64,
                        ecn,
                        sent,
                        largest_sent_pn,
                    );
                }
            } else {
                // We always start out sending ECN, so any ack that doesn't acknowledge it disables it.
                debug!("ECN not acknowledged by peer");
                self.path_data_mut(path).sending_ecn = false;
            }
        }

        self.set_loss_detection_timer(now, path);
        Ok(())
    }

    fn detect_spurious_loss(&mut self, ack: &frame::Ack, space: SpaceId, path: PathId) -> bool {
        let lost_packets = &mut self.spaces[space].for_path(path).lost_packets;

        if lost_packets.is_empty() {
            return false;
        }

        for range in ack.iter() {
            let spurious_losses: Vec<u64> = lost_packets
                .iter_range(range.clone())
                .map(|(pn, _info)| pn)
                .collect();

            for pn in spurious_losses {
                lost_packets.remove(pn);
            }
        }

        // If this ACK frame acknowledged all deemed lost packets,
        // then we have raised a spurious congestion event in the past.
        // We cannot conclude when there are remaining packets,
        // but future ACK frames might indicate a spurious loss detection.
        lost_packets.is_empty()
    }

    /// Drain lost packets that we reasonably think will never arrive
    ///
    /// The current criterion is copied from `msquic`:
    /// discard packets that were sent earlier than 2 probe timeouts ago.
    fn drain_lost_packets(&mut self, now: Instant, space: SpaceId, path: PathId) {
        let two_pto = 2 * self.path_data(path).rtt.pto_base();

        let lost_packets = &mut self.spaces[space].for_path(path).lost_packets;
        lost_packets.retain(|_pn, info| now.saturating_duration_since(info.time_sent) <= two_pto);
    }

    /// Process a new ECN block from an in-order ACK
    fn process_ecn(
        &mut self,
        now: Instant,
        space: SpaceId,
        path: PathId,
        newly_acked_pn: u64,
        ecn: frame::EcnCounts,
        largest_sent_time: Instant,
        largest_sent_pn: u64,
    ) {
        match self.spaces[space]
            .for_path(path)
            .detect_ecn(newly_acked_pn, ecn)
        {
            Err(e) => {
                debug!("halting ECN due to verification failure: {}", e);

                self.path_data_mut(path).sending_ecn = false;
                // Wipe out the existing value because it might be garbage and could interfere with
                // future attempts to use ECN on new paths.
                self.spaces[space].for_path(path).ecn_feedback = frame::EcnCounts::ZERO;
            }
            Ok(false) => {}
            Ok(true) => {
                self.path_stats.get_mut(path).congestion_events += 1;
                self.path_data_mut(path).congestion.on_congestion_event(
                    now,
                    largest_sent_time,
                    false,
                    true,
                    0,
                    largest_sent_pn,
                );
            }
        }
    }

    // Not timing-aware, so it's safe to call this for inferred acks, such as arise from
    // high-latency handshakes
    fn on_packet_acked(&mut self, now: Instant, path_id: PathId, pn: u64, info: SentPacket) {
        let path = self.path_data_mut(path_id);
        let app_limited = path.app_limited;
        path.remove_in_flight(&info);
        if info.ack_eliciting && info.path_generation == path.generation() {
            // Only pass ACKs to the congestion controller if it belongs to this exact
            // generation of the path. Otherwise we might be feeding ACKs from the previous
            // 4-tuple into our congestion controller.
            let rtt = path.rtt;
            path.congestion
                .on_ack(now, info.time_sent, info.size.into(), pn, app_limited, &rtt);
        }

        // Update state for confirmed delivery of frames
        if let Some(retransmits) = info.retransmits.get() {
            for (id, _) in retransmits.reset_stream.iter() {
                self.streams.reset_acked(*id);
            }
        }

        for frame in info.stream_frames {
            self.streams.received_ack_of(frame);
        }
    }

    fn set_key_discard_timer(&mut self, now: Instant, space: SpaceKind) {
        let start = if self.crypto_state.has_keys(EncryptionLevel::ZeroRtt) {
            now
        } else {
            self.crypto_state
                .prev_crypto
                .as_ref()
                .expect("no previous keys")
                .end_packet
                .as_ref()
                .expect("update not acknowledged yet")
                .1
        };

        // QUIC-MULTIPATH § 2.5 Key Phase Update Process: use largest PTO of all paths.
        self.timers.set(
            Timer::Conn(ConnTimer::KeyDiscard),
            start + self.max_pto_for_space(space) * 3,
            self.qlog.with_time(now),
        );
    }

    /// Handle a [`PathTimer::LossDetection`] timeout.
    ///
    /// This timer expires for two reasons:
    /// - An ACK-eliciting packet we sent should be considered lost.
    /// - The PTO may have expired and a tail-loss probe needs to be scheduled.
    ///
    /// The former needs us to schedule re-transmission of the lost data.
    ///
    /// The latter means we have not received an ACK for an ack-eliciting packet we sent
    /// within the PTO time-window. We need to schedule a tail-loss probe, an ack-eliciting
    /// packet, to try and elicit new acknowledgements. These new acknowledgements will
    /// indicate whether the previously sent packets were lost or not.
    fn on_loss_detection_timeout(&mut self, now: Instant, path_id: PathId) {
        if let Some((_, pn_space)) = self.loss_time_and_space(path_id) {
            // Time threshold loss Detection
            self.detect_lost_packets(now, pn_space, path_id, false);
            self.set_loss_detection_timer(now, path_id);
            return;
        }

        let Some((_, space)) = self.pto_time_and_space(now, path_id) else {
            debug!(%path_id, "PTO expired while unset");
            return;
        };
        trace!(
            in_flight = self.path_data(path_id).in_flight.bytes,
            count = self.path_data(path_id).pto_count,
            ?space,
            %path_id,
            "PTO fired"
        );

        let count = match self.path_data(path_id).in_flight.ack_eliciting {
            // A PTO when we're not expecting any ACKs must be due to handshake
            // anti-amplification deadlock prevention.
            0 => {
                debug_assert!(!self.peer_completed_handshake_address_validation());
                1
            }
            // Conventional loss probe
            _ => 2,
        };
        let pns = self.spaces[space].for_path(path_id);
        pns.loss_probes = pns.loss_probes.saturating_add(count);
        let path_data = self.path_data_mut(path_id);
        path_data.pto_count = path_data.pto_count.saturating_add(1);
        self.set_loss_detection_timer(now, path_id);
    }

    /// Detect any lost packets
    ///
    /// There are two cases in which we detects lost packets:
    ///
    /// - We received an ACK packet.
    /// - The [`PathTimer::LossDetection`] timer expired. So there is an un-acknowledged packet
    ///   that was followed by an acknowledged packet. The loss timer for this
    ///   un-acknowledged packet expired and we need to detect that packet as lost.
    ///
    /// Packets are lost if they are both (See RFC9002 §6.1):
    ///
    /// - Unacknowledged, in flight and sent prior to an acknowledged packet.
    /// - Old enough by either:
    ///   - Having a packet number [`TransportConfig::packet_threshold`] lower then the last
    ///     acknowledged packet.
    ///   - Being sent [`TransportConfig::time_threshold`] * RTT in the past.
    fn detect_lost_packets(
        &mut self,
        now: Instant,
        pn_space: SpaceId,
        path_id: PathId,
        due_to_ack: bool,
    ) {
        let mut lost_packets = Vec::<u64>::new();
        let mut lost_mtu_probe = None;
        let mut in_persistent_congestion = false;
        let mut size_of_lost_packets = 0u64;
        self.spaces[pn_space].for_path(path_id).loss_time = None;

        // Find all the lost packets, populating all variables initialised above.

        let path = self.path_data(path_id);
        let in_flight_mtu_probe = path.mtud.in_flight_mtu_probe();
        let loss_delay = path
            .rtt
            .conservative()
            .mul_f32(self.config.time_threshold)
            .max(TIMER_GRANULARITY);
        let first_packet_after_rtt_sample = path.first_packet_after_rtt_sample;

        let largest_acked_packet_pn = self.spaces[pn_space]
            .for_path(path_id)
            .largest_acked_packet_pn
            .expect("detect_lost_packets only to be called if path received at least one ACK");
        let packet_threshold = self.config.packet_threshold as u64;

        // InPersistentCongestion: Determine if all packets in the time period before the newest
        // lost packet, including the edges, are marked lost. PTO computation must always
        // include max ACK delay, i.e. operate as if in Data space (see RFC9001 §7.6.1).
        let congestion_period = self
            .pto(SpaceKind::Data, path_id)
            .saturating_mul(self.config.persistent_congestion_threshold);
        let mut persistent_congestion_start: Option<Instant> = None;
        let mut prev_packet = None;
        let space = self.spaces[pn_space].for_path(path_id);

        for (packet, info) in space.sent_packets.iter_range(0..largest_acked_packet_pn) {
            if prev_packet != Some(packet.wrapping_sub(1)) {
                // An intervening packet was acknowledged
                persistent_congestion_start = None;
            }

            // Packets sent before now - loss_delay are deemed lost.
            // However, we avoid subtraction as it can panic and there's no
            // saturating equivalent of this subtraction operation with a Duration.
            let packet_too_old = now.saturating_duration_since(info.time_sent) >= loss_delay;
            if packet_too_old || largest_acked_packet_pn >= packet + packet_threshold {
                // The packet should be declared lost.
                if Some(packet) == in_flight_mtu_probe {
                    // Lost MTU probes are not included in `lost_packets`, because they
                    // should not trigger a congestion control response
                    lost_mtu_probe = in_flight_mtu_probe;
                } else {
                    lost_packets.push(packet);
                    size_of_lost_packets += info.size as u64;
                    if info.ack_eliciting && due_to_ack {
                        match persistent_congestion_start {
                            // Two ACK-eliciting packets lost more than
                            // congestion_period apart, with no ACKed packets in between
                            Some(start) if info.time_sent - start > congestion_period => {
                                in_persistent_congestion = true;
                            }
                            // Persistent congestion must start after the first RTT sample
                            None if first_packet_after_rtt_sample
                                .is_some_and(|x| x < (pn_space.kind(), packet)) =>
                            {
                                persistent_congestion_start = Some(info.time_sent);
                            }
                            _ => {}
                        }
                    }
                }
            } else {
                // The packet should not yet be declared lost.
                if space.loss_time.is_none() {
                    // Since we iterate in order the lowest packet number's loss time will
                    // always be the earliest.
                    space.loss_time = Some(info.time_sent + loss_delay);
                }
                persistent_congestion_start = None;
            }

            prev_packet = Some(packet);
        }

        self.handle_lost_packets(
            pn_space,
            path_id,
            now,
            lost_packets,
            lost_mtu_probe,
            loss_delay,
            in_persistent_congestion,
            size_of_lost_packets,
        );
    }

    /// Drops the path state, declaring any remaining in-flight packets as lost
    fn discard_path(&mut self, path_id: PathId, now: Instant) {
        trace!(%path_id, "dropping path state");
        let path = self.path_data(path_id);
        let in_flight_mtu_probe = path.mtud.in_flight_mtu_probe();

        let mut size_of_lost_packets = 0u64; // add to path_stats.lost_bytes;
        let lost_pns: Vec<_> = self.spaces[SpaceId::Data]
            .for_path(path_id)
            .sent_packets
            .iter()
            .filter(|(pn, _info)| Some(*pn) != in_flight_mtu_probe)
            .map(|(pn, info)| {
                size_of_lost_packets += info.size as u64;
                pn
            })
            .collect();

        if !lost_pns.is_empty() {
            trace!(
                %path_id,
                count = lost_pns.len(),
                lost_bytes = size_of_lost_packets,
                "packets lost on path abandon"
            );
            self.handle_lost_packets(
                SpaceId::Data,
                path_id,
                now,
                lost_pns,
                in_flight_mtu_probe,
                Duration::ZERO,
                false,
                size_of_lost_packets,
            );
        }
        // Before removing the path, we fetch the final path stats via `Self::path_stats`.
        // This ensures snapshot values (like rtt) are properly updated.
        let path_stats = self.path_stats(path_id).unwrap_or_default();
        self.path_stats.discard(&path_id);
        self.partial_stats += path_stats;
        self.paths.remove(&path_id);
        self.spaces[SpaceId::Data].number_spaces.remove(&path_id);

        self.events.push_back(
            PathEvent::Discarded {
                id: path_id,
                path_stats: Box::new(path_stats),
            }
            .into(),
        );
    }

    fn handle_lost_packets(
        &mut self,
        pn_space: SpaceId,
        path_id: PathId,
        now: Instant,
        lost_packets: Vec<u64>,
        lost_mtu_probe: Option<u64>,
        loss_delay: Duration,
        in_persistent_congestion: bool,
        size_of_lost_packets: u64,
    ) {
        debug_assert!(lost_packets.is_sorted(), "lost_packets must be sorted");

        self.drain_lost_packets(now, pn_space, path_id);

        // OnPacketsLost
        if let Some(largest_lost) = lost_packets.last().cloned() {
            let old_bytes_in_flight = self.path_data_mut(path_id).in_flight.bytes;
            let largest_lost_sent = self.spaces[pn_space]
                .for_path(path_id)
                .sent_packets
                .get(largest_lost)
                .unwrap()
                .time_sent;
            let path_stats = self.path_stats.get_mut(path_id);
            path_stats.lost_packets += lost_packets.len() as u64;
            path_stats.lost_bytes += size_of_lost_packets;
            trace!(
                %path_id,
                count = lost_packets.len(),
                lost_bytes = size_of_lost_packets,
                "packets lost",
            );

            for &packet in &lost_packets {
                let Some(info) = self.spaces[pn_space].for_path(path_id).take(packet) else {
                    continue;
                };
                self.qlog
                    .emit_packet_lost(packet, &info, loss_delay, pn_space.kind(), now);
                self.paths
                    .get_mut(&path_id)
                    .unwrap()
                    .remove_in_flight(&info);

                for frame in info.stream_frames {
                    self.streams.retransmit(frame);
                }
                self.spaces[pn_space].pending |= info.retransmits;
                let path = self.path_data_mut(path_id);
                path.pending |= info.path_retransmits;
                path.mtud.on_non_probe_lost(packet, info.size);
                path.congestion.on_packet_lost(info.size, packet, now);

                self.spaces[pn_space].for_path(path_id).lost_packets.insert(
                    packet,
                    LostPacket {
                        time_sent: info.time_sent,
                    },
                );
            }

            let path = self.path_data_mut(path_id);
            if path.mtud.black_hole_detected(now) {
                path.congestion.on_mtu_update(path.mtud.current_mtu());
                if let Some(max_datagram_size) = self.datagrams().max_size()
                    && self.datagrams.drop_oversized(max_datagram_size)
                    && self.datagrams.send_blocked
                {
                    self.datagrams.send_blocked = false;
                    self.events.push_back(Event::DatagramsUnblocked);
                }
                self.path_stats.get_mut(path_id).black_holes_detected += 1;
            }

            // Don't apply congestion penalty for lost ack-only packets
            let lost_ack_eliciting =
                old_bytes_in_flight != self.path_data_mut(path_id).in_flight.bytes;

            if lost_ack_eliciting {
                self.path_stats.get_mut(path_id).congestion_events += 1;
                self.path_data_mut(path_id).congestion.on_congestion_event(
                    now,
                    largest_lost_sent,
                    in_persistent_congestion,
                    false,
                    size_of_lost_packets,
                    largest_lost,
                );
            }
        }

        // Handle a lost MTU probe
        if let Some(packet) = lost_mtu_probe {
            let info = self.spaces[SpaceId::Data]
                .for_path(path_id)
                .take(packet)
                .unwrap(); // safe: lost_mtu_probe is omitted from lost_packets, and
            // therefore must not have been removed yet
            self.paths
                .get_mut(&path_id)
                .unwrap()
                .remove_in_flight(&info);
            self.path_data_mut(path_id).mtud.on_probe_lost();
            self.path_stats.get_mut(path_id).lost_plpmtud_probes += 1;
        }
    }

    /// Returns the earliest time packets should be declared lost for all spaces on a path.
    ///
    /// If a path has an acknowledged packet with any prior un-acknowledged packets, the
    /// earliest un-acknowledged packet can be declared lost after a timeout has elapsed.
    /// The time returned is when this packet should be declared lost.
    fn loss_time_and_space(&self, path_id: PathId) -> Option<(Instant, SpaceId)> {
        SpaceId::iter()
            .filter_map(|id| {
                self.spaces[id]
                    .number_spaces
                    .get(&path_id)
                    .and_then(|pns| pns.loss_time)
                    .map(|time| (time, id))
            })
            .min_by_key(|&(time, _)| time)
    }

    /// Returns the earliest next PTO should fire for all spaces on a path.
    ///
    /// This needs to be fully deterministic because it is also used to determine the PTO
    /// that fired, not just to set the next timer. So if it fired in the past it needs to
    /// return the time from the past at which it fired.
    ///
    /// This is the next time a tail-loss probe should be sent.
    fn pto_time_and_space(&mut self, now: Instant, path_id: PathId) -> Option<(Instant, SpaceId)> {
        let path = self.path(path_id)?;
        let pto_count = path.pto_count;

        // Cap the maximum interval between two tail-loss probes.
        let max_interval = if path.rtt.get() > SLOW_RTT_THRESHOLD {
            // For slow links we want to increase the interval beyond 2s.
            (path.rtt.get() * 3) / 2
        } else if let Some(idle) = path.idle_timeout.or(self.idle_timeout)
            && idle <= MIN_IDLE_FOR_FAST_PTO
        {
            // If the idle timeout is relatively low, cap at 1s so we get plenty of retries
            // before the idle timeout fires.
            MAX_PTO_FAST_INTERVAL
        } else {
            // Otherwise cap to 2s.
            MAX_PTO_INTERVAL
        };

        if path_id == PathId::ZERO
            && path.in_flight.ack_eliciting == 0
            && !self.peer_completed_handshake_address_validation()
        {
            // Address Validation during Connection Establishment:
            // https://www.rfc-editor.org/rfc/rfc9000.html#section-8.1. To prevent a
            // deadlock if an Initial or Handshake packet from the server is lost and the
            // server can not send more due to its anti-amplification limit the client must
            // send another packet on PTO.
            let space = match self.highest_space {
                SpaceKind::Handshake => SpaceId::Handshake,
                _ => SpaceId::Initial,
            };

            let backoff = 2u32.pow(path.pto_count.min(MAX_BACKOFF_EXPONENT));
            let duration = path.rtt.pto_base() * backoff;
            let duration = duration.min(max_interval);
            return Some((now + duration, space));
        }

        let mut result = None;
        for space in SpaceId::iter() {
            let Some(pns) = self.spaces[space].number_spaces.get(&path_id) else {
                continue;
            };

            if space == SpaceId::Data && !self.is_handshake_confirmed() {
                // https://www.rfc-editor.org/rfc/rfc9002.html#section-6.2.1-7:
                // An endpoint MUST NOT set its PTO timer for the Application Data packet
                // number space until the handshake is confirmed.
                continue;
            }

            if !pns.has_in_flight() {
                continue;
            }

            // Compute the PTO duration for this space, we want to cap the maximum interval
            // between two tail-loss probes so to not do a simple exponential backoff but
            // rather iterate through the probes to compute the capped increment for an
            // exponential backoff at each step.
            let duration = {
                let max_ack_delay = if space == SpaceId::Data {
                    self.ack_frequency.max_ack_delay_for_pto()
                } else {
                    Duration::ZERO
                };
                let pto_base = path.rtt.pto_base() + max_ack_delay;
                let mut duration = pto_base;
                for i in 1..=pto_count {
                    let exponential_duration = pto_base * 2u32.pow(i.min(MAX_BACKOFF_EXPONENT));
                    let max_duration = duration + max_interval;
                    duration = exponential_duration.min(max_duration);
                }
                duration
            };

            let Some(last_ack_eliciting) = pns.time_of_last_ack_eliciting_packet else {
                continue;
            };
            // Base the deadline on when the last probe was sent, so the PTO
            // doesn't fire before the response has had time to arrive.
            let pto = last_ack_eliciting + duration;
            if result.is_none_or(|(earliest_pto, _)| pto < earliest_pto) {
                if path.anti_amplification_blocked(1) {
                    // Nothing would be able to be sent.
                    continue;
                }
                if path.in_flight.ack_eliciting == 0 {
                    // Nothing ack-eliciting, no PTO to arm/fire.
                    continue;
                }
                result = Some((pto, space));
            }
        }
        result
    }

    /// Whether the peer validated our address in the connection handshake.
    fn peer_completed_handshake_address_validation(&self) -> bool {
        if self.side.is_server() || self.state.is_closed() {
            return true;
        }
        // The server is guaranteed to have validated our address if any of our handshake or
        // 1-RTT packets are acknowledged or we've seen HANDSHAKE_DONE and discarded
        // handshake keys.
        self.spaces[SpaceId::Handshake]
            .path_space(PathId::ZERO)
            .and_then(|pns| pns.largest_acked_packet_pn)
            .is_some()
            || self.spaces[SpaceId::Data]
                .path_space(PathId::ZERO)
                .and_then(|pns| pns.largest_acked_packet_pn)
                .is_some()
            || (self.crypto_state.has_keys(EncryptionLevel::OneRtt)
                && !self.crypto_state.has_keys(EncryptionLevel::Handshake))
    }

    /// Resets the the [`PathTimer::LossDetection`] timer to the next instant it may be needed
    ///
    /// The timer must fire if either:
    /// - An ack-eliciting packet we sent needs to be declared lost.
    /// - A tail-loss probe needs to be sent.
    ///
    /// See [`Connection::on_loss_detection_timeout`] for details.
    fn set_loss_detection_timer(&mut self, now: Instant, path_id: PathId) {
        if self.state.is_closed() {
            // No loss detection takes place on closed connections, and `close_common` already
            // stopped time timer. Ensure we don't restart it inadvertently, e.g. in response to a
            // reordered packet being handled by state-insensitive code.
            return;
        }

        if let Some((loss_time, _)) = self.loss_time_and_space(path_id) {
            // Time threshold loss detection.
            self.timers.set(
                Timer::PerPath(path_id, PathTimer::LossDetection),
                loss_time,
                self.qlog.with_time(now),
            );
            return;
        }

        // Determine which PN space to arm PTO for.
        // We can only send tail-loss probes on paths that aren't abandoned yet.
        if !self.abandoned_paths.contains(&path_id)
            && let Some((timeout, _)) = self.pto_time_and_space(now, path_id)
        {
            self.timers.set(
                Timer::PerPath(path_id, PathTimer::LossDetection),
                timeout,
                self.qlog.with_time(now),
            );
        } else {
            self.timers.stop(
                Timer::PerPath(path_id, PathTimer::LossDetection),
                self.qlog.with_time(now),
            );
        }
    }

    /// The maximum probe timeout across all paths
    ///
    /// See [`Connection::pto`]
    fn max_pto_for_space(&self, space: SpaceKind) -> Duration {
        self.paths
            .keys()
            .map(|path_id| self.pto(space, *path_id))
            .max()
            .unwrap_or_else(|| {
                // No paths remain (e.g. last path was abandoned and the NoAvailablePath grace timer
                // fired before any new path was opened). Fall back to a PTO derived from the
                // configured initial RTT, matching RFC 9002 §6.2.2 initial values.
                let rtt = self.config.initial_rtt;
                let max_ack_delay = match space {
                    SpaceKind::Initial | SpaceKind::Handshake => Duration::ZERO,
                    SpaceKind::Data => self.ack_frequency.max_ack_delay_for_pto(),
                };
                rtt + cmp::max(4 * (rtt / 2), TIMER_GRANULARITY) + max_ack_delay
            })
    }

    /// Probe Timeout
    ///
    /// The PTO is logically the time in which you'd expect to receive an acknowledgement
    /// for a packet. So approximately RTT + max_ack_delay.
    fn pto(&self, space: SpaceKind, path_id: PathId) -> Duration {
        let max_ack_delay = match space {
            SpaceKind::Initial | SpaceKind::Handshake => Duration::ZERO,
            SpaceKind::Data => self.ack_frequency.max_ack_delay_for_pto(),
        };
        self.path_data(path_id).rtt.pto_base() + max_ack_delay
    }

    fn on_packet_authenticated(
        &mut self,
        now: Instant,
        space_id: SpaceKind,
        path_id: PathId,
        ecn: Option<EcnCodepoint>,
        packet_number: Option<u64>,
        spin: bool,
        is_1rtt: bool,
        remote: &FourTuple,
    ) {
        // During the handshake we already have discarded packets that do not match the path
        // remote. So any off-path packet here is either a probing packet or a
        // migration. Handling probing packets here means that the path's idle timeout will
        // be reset and will delay detecting the path as idle. However tail-loss probes
        // would still not get acknowledged if the path was broken so eventually the path
        // would still become idle.
        let is_on_path = self
            .path_data(path_id)
            .network_path
            .is_probably_same_path(remote);

        self.total_authed_packets += 1;
        self.reset_keep_alive(path_id, now);
        self.reset_idle_timeout(now, space_id, path_id);
        self.path_data_mut(path_id).permit_idle_reset = true;

        // Do not process ECN for off-path packets. If this is a migration we'll get ECN
        // back once we've migrated.
        if is_on_path {
            self.receiving_ecn |= ecn.is_some();
            if let Some(x) = ecn {
                let space = &mut self.spaces[space_id];
                space.for_path(path_id).ecn_counters += x;

                if x.is_ce() {
                    space
                        .for_path(path_id)
                        .pending_acks
                        .set_immediate_ack_required();
                }
            }
        }

        let Some(packet_number) = packet_number else {
            return;
        };
        match &self.side {
            ConnectionSide::Client { .. } => {
                // If we received a handshake packet that authenticated, then we're talking to
                // the real server.  From now on we should no longer allow the server to migrate
                // its address.
                if space_id == SpaceKind::Handshake
                    && let Some(hs) = self.state.as_handshake_mut()
                {
                    hs.allow_server_migration = false;
                }
            }
            ConnectionSide::Server { .. } => {
                if self.crypto_state.has_keys(EncryptionLevel::Initial)
                    && space_id == SpaceKind::Handshake
                {
                    // A server stops sending and processing Initial packets when it receives its first Handshake packet.
                    self.discard_space(now, SpaceKind::Initial);
                }
                if self.crypto_state.has_keys(EncryptionLevel::ZeroRtt) && is_1rtt {
                    // Discard 0-RTT keys soon after receiving a 1-RTT packet
                    self.set_key_discard_timer(now, space_id)
                }
            }
        }
        let space = self.spaces[space_id].for_path(path_id);

        space.pending_acks.insert_one(packet_number, now);
        if packet_number >= space.largest_received_packet_number.unwrap_or_default() {
            space.largest_received_packet_number = Some(packet_number);

            // Update outgoing spin bit for on-path packets, inverting iff we're the client
            if is_on_path {
                self.spin = self.side.is_client() ^ spin;
            }
        }
    }

    /// Resets the idle timeout timers
    ///
    /// Without multipath there is only the connection-wide idle timeout. When multipath is
    /// enabled there is an additional per-path idle timeout.
    fn reset_idle_timeout(&mut self, now: Instant, space: SpaceKind, path_id: PathId) {
        // First reset the global idle timeout.
        if let Some(timeout) = self.idle_timeout {
            if self.state.is_closed() {
                self.timers
                    .stop(Timer::Conn(ConnTimer::Idle), self.qlog.with_time(now));
            } else {
                let dt = cmp::max(timeout, 3 * self.max_pto_for_space(space));
                self.timers.set(
                    Timer::Conn(ConnTimer::Idle),
                    now + dt,
                    self.qlog.with_time(now),
                );
            }
        }

        // Now handle the per-path state
        if let Some(timeout) = self.path_data(path_id).idle_timeout {
            if self.state.is_closed() {
                self.timers.stop(
                    Timer::PerPath(path_id, PathTimer::PathIdle),
                    self.qlog.with_time(now),
                );
            } else {
                let dt = cmp::max(timeout, 3 * self.pto(space, path_id));
                self.timers.set(
                    Timer::PerPath(path_id, PathTimer::PathIdle),
                    now + dt,
                    self.qlog.with_time(now),
                );
            }
        }
    }

    /// Resets both the [`ConnTimer::KeepAlive`] and [`PathTimer::PathKeepAlive`] timers
    fn reset_keep_alive(&mut self, path_id: PathId, now: Instant) {
        if !self.state.is_established() {
            return;
        }

        if let Some(interval) = self.config.keep_alive_interval {
            self.timers.set(
                Timer::Conn(ConnTimer::KeepAlive),
                now + interval,
                self.qlog.with_time(now),
            );
        }

        if let Some(interval) = self.path_data(path_id).keep_alive {
            self.timers.set(
                Timer::PerPath(path_id, PathTimer::PathKeepAlive),
                now + interval,
                self.qlog.with_time(now),
            );
        }
    }

    /// Sets the timer for when a previously issued CID should be retired next
    fn reset_cid_retirement(&mut self, now: Instant) {
        if let Some((_path, t)) = self.next_cid_retirement() {
            self.timers.set(
                Timer::Conn(ConnTimer::PushNewCid),
                t,
                self.qlog.with_time(now),
            );
        }
    }

    /// The next time when a previously issued CID should be retired
    fn next_cid_retirement(&self) -> Option<(PathId, Instant)> {
        self.local_cid_state
            .iter()
            .filter_map(|(path_id, cid_state)| cid_state.next_timeout().map(|t| (*path_id, t)))
            .min_by_key(|(_path_id, timeout)| *timeout)
    }

    /// Handle the already-decrypted first packet from the client
    ///
    /// Decrypting the first packet in the `Endpoint` allows stateless packet handling to be more
    /// efficient.
    pub(crate) fn handle_first_packet(
        &mut self,
        now: Instant,
        network_path: FourTuple,
        ecn: Option<EcnCodepoint>,
        packet_number: u64,
        packet: InitialPacket,
        remaining: Option<BytesMut>,
    ) -> Result<(), ConnectionError> {
        let span = trace_span!("first recv");
        let _guard = span.enter();
        debug_assert!(self.side.is_server());
        let len = packet.header_data.len() + packet.payload.len();
        let path_id = PathId::ZERO;
        self.path_data_mut(path_id).total_recvd = len as u64;

        if let Some(hs) = self.state.as_handshake_mut() {
            hs.expected_token = packet.header.token.clone();
        } else {
            unreachable!("first packet must be delivered in Handshake state");
        }

        // The first packet is always on PathId::ZERO
        self.on_packet_authenticated(
            now,
            SpaceKind::Initial,
            path_id,
            ecn,
            Some(packet_number),
            false,
            false,
            &network_path,
        );

        let packet: Packet = packet.into();

        let mut qlog = QlogRecvPacket::new(len);
        qlog.header(&packet.header, Some(packet_number), path_id);

        self.process_decrypted_packet(
            now,
            network_path,
            path_id,
            Some(packet_number),
            packet,
            &mut qlog,
        )?;
        self.qlog.emit_packet_received(qlog, now);
        if let Some(data) = remaining {
            self.handle_coalesced(now, network_path, path_id, ecn, data);
        }

        self.qlog.emit_recovery_metrics(
            path_id,
            &mut self
                .paths
                .get_mut(&path_id)
                .expect("path_id was supplied by the caller for an active path")
                .data,
            now,
        );

        Ok(())
    }

    fn init_0rtt(&mut self, now: Instant) {
        let Some((header, packet)) = self.crypto_state.session.early_crypto() else {
            return;
        };
        if self.side.is_client() {
            match self.crypto_state.session.transport_parameters() {
                Ok(params) => {
                    let params = params
                        .expect("crypto layer didn't supply transport parameters with ticket");
                    // Certain values must not be cached
                    let params = TransportParameters {
                        initial_src_cid: None,
                        original_dst_cid: None,
                        preferred_address: None,
                        retry_src_cid: None,
                        stateless_reset_token: None,
                        min_ack_delay: None,
                        ack_delay_exponent: TransportParameters::default().ack_delay_exponent,
                        max_ack_delay: TransportParameters::default().max_ack_delay,
                        initial_max_path_id: None,
                        ..params
                    };
                    self.set_peer_params(params);
                    self.qlog.emit_peer_transport_params_restored(self, now);
                }
                Err(e) => {
                    error!("session ticket has malformed transport parameters: {}", e);
                    return;
                }
            }
        }
        trace!("0-RTT enabled");
        self.crypto_state.enable_zero_rtt(header, packet);
    }

    fn read_crypto(
        &mut self,
        space: SpaceId,
        crypto: &frame::Crypto,
        payload_len: usize,
    ) -> Result<(), TransportError> {
        let expected = if !self.state.is_handshake() {
            SpaceId::Data
        } else if self.highest_space == SpaceKind::Initial {
            SpaceId::Initial
        } else {
            // On the server, self.highest_space can be Data after receiving the client's first
            // flight, but we expect Handshake CRYPTO until the handshake is complete.
            SpaceId::Handshake
        };
        // We can't decrypt Handshake packets when highest_space is Initial, CRYPTO frames in 0-RTT
        // packets are illegal, and we don't process 1-RTT packets until the handshake is
        // complete. Therefore, we will never see CRYPTO data from a later-than-expected space.
        debug_assert!(space <= expected, "received out-of-order CRYPTO data");

        let end = crypto.offset + crypto.data.len() as u64;
        if space < expected
            && end
                > self.crypto_state.spaces[space.kind()]
                    .crypto_stream
                    .bytes_read()
        {
            warn!(
                "received new {:?} CRYPTO data when expecting {:?}",
                space, expected
            );
            return Err(TransportError::PROTOCOL_VIOLATION(
                "new data at unexpected encryption level",
            ));
        }

        let crypto_space = &mut self.crypto_state.spaces[space.kind()];
        let max = end.saturating_sub(crypto_space.crypto_stream.bytes_read());
        if max > self.config.crypto_buffer_size as u64 {
            return Err(TransportError::CRYPTO_BUFFER_EXCEEDED(""));
        }

        crypto_space
            .crypto_stream
            .insert(crypto.offset, crypto.data.clone(), payload_len);
        while let Some(chunk) = crypto_space.crypto_stream.read(usize::MAX, true) {
            trace!("consumed {} CRYPTO bytes", chunk.bytes.len());
            if self.crypto_state.session.read_handshake(&chunk.bytes)? {
                self.events.push_back(Event::HandshakeDataReady);
            }
        }

        Ok(())
    }

    fn write_crypto(&mut self) {
        loop {
            let space = self.highest_space;
            let mut outgoing = Vec::new();
            if let Some(crypto) = self.crypto_state.session.write_handshake(&mut outgoing) {
                match space {
                    SpaceKind::Initial => {
                        self.upgrade_crypto(SpaceKind::Handshake, crypto);
                    }
                    SpaceKind::Handshake => {
                        self.upgrade_crypto(SpaceKind::Data, crypto);
                    }
                    SpaceKind::Data => unreachable!("got updated secrets during 1-RTT"),
                }
            }
            if outgoing.is_empty() {
                if space == self.highest_space {
                    break;
                } else {
                    // Keys updated, check for more data to send
                    continue;
                }
            }
            let offset = self.crypto_state.spaces[space].crypto_offset;
            let outgoing = Bytes::from(outgoing);
            if let Some(hs) = self.state.as_handshake_mut()
                && space == SpaceKind::Initial
                && offset == 0
                && self.side.is_client()
            {
                hs.client_hello = Some(outgoing.clone());
            }
            self.crypto_state.spaces[space].crypto_offset += outgoing.len() as u64;
            trace!("wrote {} {:?} CRYPTO bytes", outgoing.len(), space);
            self.spaces[space].pending.crypto.push_back(frame::Crypto {
                offset,
                data: outgoing,
            });
        }
    }

    /// Switch to stronger cryptography during handshake
    fn upgrade_crypto(&mut self, space: SpaceKind, crypto: Keys) {
        debug_assert!(
            !self.crypto_state.has_keys(space.encryption_level()),
            "already reached packet space {space:?}"
        );
        trace!("{:?} keys ready", space);
        if space == SpaceKind::Data {
            // Precompute the first key update
            self.crypto_state.next_crypto = Some(
                self.crypto_state
                    .session
                    .next_1rtt_keys()
                    .expect("handshake should be complete"),
            );
        }

        self.crypto_state.spaces[space].keys = Some(crypto);
        debug_assert!(space > self.highest_space);
        self.highest_space = space;
        if space == SpaceKind::Data && self.side.is_client() {
            // Discard 0-RTT keys because 1-RTT keys are available.
            self.crypto_state.discard_zero_rtt();
        }
    }

    fn discard_space(&mut self, now: Instant, space: SpaceKind) {
        debug_assert!(space != SpaceKind::Data);
        trace!("discarding {:?} keys", space);
        if space == SpaceKind::Initial {
            // No longer needed
            if let ConnectionSide::Client { token, .. } = &mut self.side {
                *token = Bytes::new();
            }
        }
        self.crypto_state.spaces[space].keys = None;
        let space = &mut self.spaces[space];
        let pns = space.for_path(PathId::ZERO);
        pns.time_of_last_ack_eliciting_packet = None;
        pns.loss_time = None;
        pns.loss_probes = 0;
        let sent_packets = mem::take(&mut pns.sent_packets);
        let path = self
            .paths
            .get_mut(&PathId::ZERO)
            .expect("PathId::ZERO is alive while Initial/Handshake spaces exist");
        for (_, packet) in sent_packets.into_iter() {
            path.data.remove_in_flight(&packet);
        }

        self.set_loss_detection_timer(now, PathId::ZERO)
    }

    fn handle_coalesced(
        &mut self,
        now: Instant,
        network_path: FourTuple,
        path_id: PathId,
        ecn: Option<EcnCodepoint>,
        data: BytesMut,
    ) {
        self.path_data_mut(path_id)
            .inc_total_recvd(data.len() as u64);
        let mut remaining = Some(data);
        let cid_len = self
            .local_cid_state
            .values()
            .map(|cid_state| cid_state.cid_len())
            .next()
            .expect("one cid_state must exist");
        while let Some(data) = remaining {
            match PartialDecode::new(
                data,
                &FixedLengthConnectionIdParser::new(cid_len),
                &[self.version],
                self.endpoint_config.grease_quic_bit,
            ) {
                Ok((partial_decode, rest)) => {
                    remaining = rest;
                    self.handle_decode(now, network_path, path_id, ecn, partial_decode);
                }
                Err(e) => {
                    trace!("malformed header: {}", e);
                    return;
                }
            }
        }
    }

    /// Decrypts the packet and processes the payload.
    ///
    /// Processes the entire packet, starting with removing header protection, then handling
    /// a stateless reset if needed, and decrypting and processing the frames in the payload
    /// if not a stateless reset.
    fn handle_decode(
        &mut self,
        now: Instant,
        network_path: FourTuple,
        path_id: PathId,
        ecn: Option<EcnCodepoint>,
        partial_decode: PartialDecode,
    ) {
        let qlog = QlogRecvPacket::new(partial_decode.len());
        if let Some(decoded) = self
            .crypto_state
            .unprotect_header(partial_decode, self.peer_params.stateless_reset_token)
        {
            self.handle_packet(
                now,
                network_path,
                path_id,
                ecn,
                decoded.packet,
                decoded.stateless_reset,
                qlog,
            );
        }
    }

    /// Handles a packet with header protection removed.
    ///
    /// The packet body is still encrypted at this point.
    ///
    /// If the datagram was a stateless reset we may have failed to remove header protection
    /// and thus `packet` may be `None`.
    fn handle_packet(
        &mut self,
        now: Instant,
        network_path: FourTuple,
        path_id: PathId,
        ecn: Option<EcnCodepoint>,
        packet: Option<Packet>,
        stateless_reset: bool,
        mut qlog: QlogRecvPacket,
    ) {
        if let Some(ref packet) = packet {
            trace!(
                "got {:?} packet ({} bytes) from {} using id {}",
                packet.header.space(),
                packet.payload.len() + packet.header_data.len(),
                network_path,
                packet.header.dst_cid(),
            );
        }

        let was_closed = self.state.is_closed();
        let was_drained = self.state.is_drained();

        // Now decrypt the packet payload in-place.
        let decrypted = match packet {
            None => Err(None),
            Some(mut packet) => self
                .decrypt_packet(now, path_id, &mut packet)
                .map(move |number| (packet, number)),
        };
        let result = match decrypted {
            _ if stateless_reset => {
                debug!("got stateless reset");
                Err(ConnectionError::Reset)
            }
            Err(Some(e)) => {
                warn!("illegal packet: {}", e);
                Err(e.into())
            }
            Err(None) => {
                debug!("failed to authenticate packet");
                self.authentication_failures += 1;
                let integrity_limit = self
                    .crypto_state
                    .integrity_limit(self.highest_space)
                    .unwrap();
                if self.authentication_failures > integrity_limit {
                    Err(TransportError::AEAD_LIMIT_REACHED("integrity limit violated").into())
                } else {
                    return;
                }
            }
            Ok((packet, pn)) => {
                // We received an authenticated packet and decrypted it.
                qlog.header(&packet.header, pn, path_id);
                let span = match pn {
                    Some(pn) => trace_span!("recv", space = ?packet.header.space(), pn),
                    None => trace_span!("recv", space = ?packet.header.space()),
                };
                let _guard = span.enter();

                // Now the packet is authenticated we do the migration during the handshake,
                // see Handshake::allow_server_migration for details.  Be careful here to
                // not yet rely on the path existing however, new paths are accepted and
                // created later.
                // Note that we can't do any other migrations yet, for those we need to know
                // whether this was a probing packet or not. See the end of
                // Self::process_packet for that.
                if self.is_handshaking()
                    && self
                        .path(path_id)
                        .map(|path_data| {
                            !path_data.network_path.is_probably_same_path(&network_path)
                        })
                        .unwrap_or(false)
                {
                    if let Some(hs) = self.state.as_handshake()
                        && hs.allow_server_migration
                    {
                        trace!(
                            %network_path,
                            prev = %self.path_data(path_id).network_path,
                            "server migrated to new remote",
                        );
                        self.path_data_mut(path_id).network_path = network_path;
                        self.qlog.emit_tuple_assigned(path_id, network_path, now);
                    } else {
                        debug!(
                            recv_path = %network_path,
                            expected_path = %self.path_data_mut(path_id).network_path,
                            "discarding packet with unexpected remote during handshake",
                        );
                        return;
                    }
                }

                let dedup = self.spaces[packet.header.space()]
                    .path_space_mut(path_id)
                    .map(|pns| &mut pns.dedup);
                if pn.zip(dedup).is_some_and(|(n, d)| d.insert(n)) {
                    debug!("discarding possible duplicate packet");
                    self.qlog.emit_packet_received(qlog, now);
                    return;
                } else if self.state.is_handshake() && packet.header.is_short() {
                    // TODO: SHOULD buffer these to improve reordering tolerance.
                    trace!("dropping short packet during handshake");
                    self.qlog.emit_packet_received(qlog, now);
                    return;
                } else {
                    if let Header::Initial(InitialHeader { ref token, .. }) = packet.header
                        && let Some(hs) = self.state.as_handshake()
                        && self.side.is_server()
                        && token != &hs.expected_token
                    {
                        // Clients must send the same retry token in every Initial. Initial
                        // packets can be spoofed, so we discard rather than killing the
                        // connection.
                        warn!("discarding Initial with invalid retry token");
                        self.qlog.emit_packet_received(qlog, now);
                        return;
                    }

                    if !self.state.is_closed() {
                        let spin = match packet.header {
                            Header::Short { spin, .. } => spin,
                            _ => false,
                        };

                        if self.side().is_server() && !self.abandoned_paths.contains(&path_id) {
                            // Only the client is allowed to open paths
                            self.create_path(path_id, network_path, now, pn);
                        }
                        if self.paths.contains_key(&path_id) {
                            self.on_packet_authenticated(
                                now,
                                packet.header.space(),
                                path_id,
                                ecn,
                                pn,
                                spin,
                                packet.header.is_1rtt(),
                                &network_path,
                            );
                        }
                    }

                    let res = self.process_decrypted_packet(
                        now,
                        network_path,
                        path_id,
                        pn,
                        packet,
                        &mut qlog,
                    );

                    self.qlog.emit_packet_received(qlog, now);
                    res
                }
            }
        };

        // State transitions for error cases
        if let Err(conn_err) = result {
            match conn_err {
                ConnectionError::ApplicationClosed(reason) => self.state.move_to_closed(reason),
                ConnectionError::ConnectionClosed(reason) => self.state.move_to_closed(reason),
                ConnectionError::Reset
                | ConnectionError::TransportError(TransportError {
                    code: TransportErrorCode::AEAD_LIMIT_REACHED,
                    ..
                }) => {
                    if !self.state.is_drained() {
                        self.state
                            .move_to_drained(Some(conn_err), &mut self.endpoint_events);
                    }
                }
                ConnectionError::TimedOut => {
                    unreachable!("timeouts aren't generated by packet processing");
                }
                ConnectionError::TransportError(err) => {
                    debug!("closing connection due to transport error: {}", err);
                    self.state.move_to_closed(err);
                }
                ConnectionError::VersionMismatch => {
                    self.state
                        .move_to_draining(Some(conn_err), &mut self.endpoint_events);
                }
                ConnectionError::LocallyClosed => {
                    unreachable!("LocallyClosed isn't generated by packet processing");
                }
                ConnectionError::CidsExhausted => {
                    unreachable!("CidsExhausted isn't generated by packet processing");
                }
            };
        }

        if !was_closed && self.state.is_closed() {
            self.close_common();
            if !self.state.is_drained() {
                self.set_close_timer(now);
            }
        }
        if !was_drained && self.state.is_drained() {
            // Close timer may have been started previously, e.g. if we sent a close and got a
            // stateless reset in response
            self.timers
                .stop(Timer::Conn(ConnTimer::Close), self.qlog.with_time(now));
        }

        // Transmit CONNECTION_CLOSE if necessary.
        //
        // If we received a valid packet and we are in the closed state we should respond
        // with a CONNECTION_CLOSE frame.
        // TODO: This SHOULD be rate-limited according to §10.2.1 of QUIC-TRANSPORT, but
        //    that does not yet happen. This is triggered by each received packet.
        if matches!(self.state.as_type(), StateType::Closed) {
            // From https://www.rfc-editor.org/rfc/rfc9000.html#section-10.2.1-7
            //
            // While in the closing state we must either:
            // - discard packets coming from an un-validated remote OR
            // - ensure we do not send more than 3 times the received data
            //
            // Doing the 2nd would mean we would be able to send CONNECTION_CLOSE to a peer
            // who was (involuntary) migrated just at the time we initiated immediate
            // close. It is a lot more work though. So while we would like to do this for
            // now we only do 1.
            //
            // Another shortcoming of the current implementation is that when we have a
            // previous PathData which is validated and the remote matches that path, we
            // should schedule CONNECTION_CLOSE on that path. However currently we can not
            // schedule such a packet. We should also fix this some day. This makes us
            // vulnerable to an attacker faking a migration at the right time and then we'd
            // be unable to send the CONNECTION_CLOSE to the real remote.
            if self
                .paths
                .get(&path_id)
                .map(|p| p.data.validated && p.data.network_path == network_path)
                .unwrap_or(false)
            {
                self.connection_close_pending = true;
            }
        }
    }

    fn process_decrypted_packet(
        &mut self,
        now: Instant,
        network_path: FourTuple,
        path_id: PathId,
        number: Option<u64>,
        packet: Packet,
        qlog: &mut QlogRecvPacket,
    ) -> Result<(), ConnectionError> {
        if !self.paths.contains_key(&path_id) {
            // There is a chance this is a server side, first (for this path) packet, which would
            // be a protocol violation. It's more likely, however, that this is a packet of a
            // pruned path
            trace!(%path_id, ?number, "discarding packet for unknown path");
            return Ok(());
        }
        let state = match self.state.as_type() {
            StateType::Established => {
                match packet.header.space() {
                    SpaceKind::Data => self.process_payload(
                        now,
                        network_path,
                        path_id,
                        number.unwrap(),
                        packet,
                        qlog,
                    )?,
                    _ if packet.header.has_frames() => {
                        self.process_early_payload(now, path_id, packet, qlog)?
                    }
                    _ => {
                        trace!("discarding unexpected pre-handshake packet");
                    }
                }
                return Ok(());
            }
            StateType::Closed => {
                for result in frame::Iter::new(packet.payload.freeze())? {
                    let frame = match result {
                        Ok(frame) => frame,
                        Err(err) => {
                            debug!("frame decoding error: {err:?}");
                            continue;
                        }
                    };
                    qlog.frame(&frame);

                    if let Frame::Padding = frame {
                        continue;
                    };

                    trace!(?frame, "processing frame in closed state");

                    self.path_stats.get_mut(path_id).frame_rx.record(frame.ty());

                    if let Frame::Close(_error) = frame {
                        self.state.move_to_draining(None, &mut self.endpoint_events);
                        break;
                    }
                }
                return Ok(());
            }
            StateType::Draining | StateType::Drained => return Ok(()),
            StateType::Handshake => self.state.as_handshake_mut().expect("checked"),
        };

        match packet.header {
            Header::Retry {
                src_cid: remote_cid,
                ..
            } => {
                debug_assert_eq!(path_id, PathId::ZERO);
                if self.side.is_server() {
                    return Err(TransportError::PROTOCOL_VIOLATION("client sent Retry").into());
                }

                let is_valid_retry = self
                    .remote_cids
                    .get(&path_id)
                    .map(|cids| cids.active())
                    .map(|orig_dst_cid| {
                        self.crypto_state.session.is_valid_retry(
                            orig_dst_cid,
                            &packet.header_data,
                            &packet.payload,
                        )
                    })
                    .unwrap_or_default();
                if self.total_authed_packets > 1
                            || packet.payload.len() <= 16 // token + 16 byte tag
                            || !is_valid_retry
                {
                    trace!("discarding invalid Retry");
                    // - After the client has received and processed an Initial or Retry
                    //   packet from the server, it MUST discard any subsequent Retry
                    //   packets that it receives.
                    // - A client MUST discard a Retry packet with a zero-length Retry Token
                    //   field.
                    // - Clients MUST discard Retry packets that have a Retry Integrity Tag
                    //   that cannot be validated
                    return Ok(());
                }

                trace!("retrying with CID {}", remote_cid);
                let client_hello = state.client_hello.take().unwrap();
                self.retry_src_cid = Some(remote_cid);
                self.remote_cids
                    .get_mut(&path_id)
                    .expect("PathId::ZERO not yet abandoned, is_valid_retry would have been false")
                    .update_initial_cid(remote_cid);
                self.remote_handshake_cid = remote_cid;

                let space = &mut self.spaces[SpaceId::Initial];
                if let Some(info) = space.for_path(PathId::ZERO).take(0) {
                    self.on_packet_acked(now, PathId::ZERO, 0, info);
                };

                self.discard_space(now, SpaceKind::Initial); // Make sure we clean up after
                // any retransmitted Initials
                let crypto_space = &mut self.crypto_state.spaces[SpaceKind::Initial];
                crypto_space.keys = Some(
                    self.crypto_state
                        .session
                        .initial_keys(remote_cid, self.side.side()),
                );
                crypto_space.crypto_offset = client_hello.len() as u64;

                let next_pn = self.spaces[SpaceId::Initial]
                    .for_path(path_id)
                    .next_packet_number;
                self.spaces[SpaceId::Initial] = {
                    let mut space = PacketSpace::new(now, SpaceId::Initial, &mut self.rng);
                    space.for_path(path_id).next_packet_number = next_pn;
                    space.pending.crypto.push_back(frame::Crypto {
                        offset: 0,
                        data: client_hello,
                    });
                    space
                };

                // Retransmit all 0-RTT data
                let zero_rtt = mem::take(
                    &mut self.spaces[SpaceId::Data]
                        .for_path(PathId::ZERO)
                        .sent_packets,
                );
                for (_, info) in zero_rtt.into_iter() {
                    self.paths
                        .get_mut(&PathId::ZERO)
                        .unwrap()
                        .remove_in_flight(&info);
                    self.spaces[SpaceId::Data].pending |= info.retransmits;
                }
                self.streams.retransmit_all_for_0rtt();

                let token_len = packet.payload.len() - 16;
                let ConnectionSide::Client { ref mut token, .. } = self.side else {
                    unreachable!("we already short-circuited if we're server");
                };
                *token = packet.payload.freeze().split_to(token_len);

                self.state = State::handshake(state::Handshake {
                    expected_token: Bytes::new(),
                    remote_cid_set: false,
                    client_hello: None,
                    allow_server_migration: self.config.server_handshake_migration,
                });
                Ok(())
            }
            Header::Long {
                ty: LongType::Handshake,
                src_cid: remote_cid,
                dst_cid: local_cid,
                ..
            } => {
                debug_assert_eq!(path_id, PathId::ZERO);
                if remote_cid != self.remote_handshake_cid {
                    debug!(
                        "discarding packet with mismatched remote CID: {} != {}",
                        self.remote_handshake_cid, remote_cid
                    );
                    return Ok(());
                }
                self.on_path_validated(path_id);

                self.process_early_payload(now, path_id, packet, qlog)?;
                if self.state.is_closed() {
                    return Ok(());
                }

                if self.crypto_state.session.is_handshaking() {
                    trace!("handshake ongoing");
                    return Ok(());
                }

                if self.side.is_client() {
                    // Client-only because server params were set from the client's Initial
                    let params = self
                        .crypto_state
                        .session
                        .transport_parameters()?
                        .ok_or_else(|| {
                            TransportError::new(
                                TransportErrorCode::crypto(0x6d),
                                "transport parameters missing".to_owned(),
                            )
                        })?;

                    if self.has_0rtt() {
                        if !self.crypto_state.session.early_data_accepted().unwrap() {
                            debug_assert!(self.side.is_client());
                            debug!("0-RTT rejected");
                            self.crypto_state.accepted_0rtt = false;
                            self.streams.zero_rtt_rejected();

                            // Discard already-queued frames
                            self.spaces[SpaceId::Data].pending = Retransmits::default();

                            // Discard 0-RTT packets
                            let sent_packets = mem::take(
                                &mut self.spaces[SpaceId::Data].for_path(path_id).sent_packets,
                            );
                            for (_, packet) in sent_packets.into_iter() {
                                self.paths
                                    .get_mut(&path_id)
                                    .unwrap()
                                    .remove_in_flight(&packet);
                            }
                        } else {
                            self.crypto_state.accepted_0rtt = true;
                            params.validate_resumption_from(&self.peer_params)?;
                        }
                    }
                    if let Some(token) = params.stateless_reset_token {
                        let remote = self.path_data(path_id).network_path.remote;
                        debug_assert!(!self.state.is_drained()); // requirement for endpoint events, checked above
                        self.endpoint_events
                            .push_back(EndpointEventInner::ResetToken(path_id, remote, token));
                    }
                    self.handle_peer_params(params, local_cid, remote_cid, now)?;
                    self.issue_first_cids(now);
                } else {
                    // Server-only
                    self.spaces[SpaceId::Data].pending.handshake_done = true;
                    self.discard_space(now, SpaceKind::Handshake);
                    self.events.push_back(Event::HandshakeConfirmed);
                    trace!("handshake confirmed");
                }

                self.events.push_back(Event::Connected);
                self.state.move_to_established();
                trace!("established");

                // Multipath can only be enabled after the state has reached Established.
                // So this can not happen any earlier.
                self.issue_first_path_cids(now);
                Ok(())
            }
            Header::Initial(InitialHeader {
                src_cid: remote_cid,
                dst_cid: local_cid,
                ..
            }) => {
                debug_assert_eq!(path_id, PathId::ZERO);
                if !state.remote_cid_set {
                    trace!("switching remote CID to {}", remote_cid);
                    let mut state = state.clone();
                    self.remote_cids
                        .get_mut(&path_id)
                        .expect("PathId::ZERO not yet abandoned")
                        .update_initial_cid(remote_cid);
                    self.remote_handshake_cid = remote_cid;
                    self.original_remote_cid = remote_cid;
                    state.remote_cid_set = true;
                    self.state.move_to_handshake(state);
                } else if remote_cid != self.remote_handshake_cid {
                    debug!(
                        "discarding packet with mismatched remote CID: {} != {}",
                        self.remote_handshake_cid, remote_cid
                    );
                    return Ok(());
                }

                let starting_space = self.highest_space;
                self.process_early_payload(now, path_id, packet, qlog)?;

                if self.side.is_server()
                    && starting_space == SpaceKind::Initial
                    && self.highest_space != SpaceKind::Initial
                {
                    let params = self
                        .crypto_state
                        .session
                        .transport_parameters()?
                        .ok_or_else(|| {
                            TransportError::new(
                                TransportErrorCode::crypto(0x6d),
                                "transport parameters missing".to_owned(),
                            )
                        })?;
                    self.handle_peer_params(params, local_cid, remote_cid, now)?;
                    self.issue_first_cids(now);
                    self.init_0rtt(now);
                }
                Ok(())
            }
            Header::Long {
                ty: LongType::ZeroRtt,
                ..
            } => {
                self.process_payload(now, network_path, path_id, number.unwrap(), packet, qlog)?;
                Ok(())
            }
            Header::VersionNegotiate { .. } => {
                if self.total_authed_packets > 1 {
                    return Ok(());
                }
                let supported = packet
                    .payload
                    .chunks(4)
                    .any(|x| match <[u8; 4]>::try_from(x) {
                        Ok(version) => self.version == u32::from_be_bytes(version),
                        Err(_) => false,
                    });
                if supported {
                    return Ok(());
                }
                debug!("remote doesn't support our version");
                Err(ConnectionError::VersionMismatch)
            }
            Header::Short { .. } => unreachable!(
                "short packets received during handshake are discarded in handle_packet"
            ),
        }
    }

    /// Process an Initial or Handshake packet payload
    fn process_early_payload(
        &mut self,
        now: Instant,
        path_id: PathId,
        packet: Packet,
        #[allow(unused)] qlog: &mut QlogRecvPacket,
    ) -> Result<(), TransportError> {
        debug_assert_ne!(packet.header.space(), SpaceKind::Data);
        debug_assert_eq!(path_id, PathId::ZERO);
        let payload_len = packet.payload.len();
        let mut ack_eliciting = false;
        for result in frame::Iter::new(packet.payload.freeze())? {
            let frame = result?;
            qlog.frame(&frame);
            let span = match frame {
                Frame::Padding => continue,
                _ => Some(trace_span!("frame", ty = %frame.ty(), path = tracing::field::Empty)),
            };

            self.path_stats.get_mut(path_id).frame_rx.record(frame.ty());

            let _guard = span.as_ref().map(|x| x.enter());
            ack_eliciting |= frame.is_ack_eliciting();

            // Process frames
            if frame.is_1rtt() && packet.header.space() != SpaceKind::Data {
                return Err(TransportError::PROTOCOL_VIOLATION(
                    "illegal frame type in handshake",
                ));
            }

            match frame {
                Frame::Padding | Frame::Ping => {}
                Frame::Crypto(frame) => {
                    self.read_crypto(packet.header.space().into(), &frame, payload_len)?;
                }
                Frame::Ack(ack) => {
                    self.on_ack_received(now, packet.header.space().into(), ack)?;
                }
                Frame::PathAck(ack) => {
                    span.as_ref()
                        .map(|span| span.record("path", tracing::field::display(&ack.path_id)));
                    self.on_path_ack_received(now, packet.header.space().into(), ack)?;
                }
                Frame::Close(reason) => {
                    self.state
                        .move_to_draining(Some(reason.into()), &mut self.endpoint_events);
                    return Ok(());
                }
                _ => {
                    let mut err =
                        TransportError::PROTOCOL_VIOLATION("illegal frame type in handshake");
                    err.frame = frame::MaybeFrame::Known(frame.ty());
                    return Err(err);
                }
            }
        }

        if ack_eliciting {
            // In the initial and handshake spaces, ACKs must be sent immediately
            self.spaces[packet.header.space()]
                .for_path(path_id)
                .pending_acks
                .set_immediate_ack_required();
        }

        self.write_crypto();
        Ok(())
    }

    /// Processes the decrypted packet payload, always in the data space.
    fn process_payload(
        &mut self,
        now: Instant,
        network_path: FourTuple,
        path_id: PathId,
        number: u64,
        packet: Packet,
        #[allow(unused)] qlog: &mut QlogRecvPacket,
    ) -> Result<(), TransportError> {
        let payload = packet.payload.freeze();
        let mut is_probing_packet = true;
        let mut close = None;
        let payload_len = payload.len();
        let mut ack_eliciting = false;
        // if this packet triggers a path migration and includes a observed address frame, it's
        // stored here
        let mut migration_observed_addr = None;
        for result in frame::Iter::new(payload)? {
            let frame = result?;
            qlog.frame(&frame);
            let span = match frame {
                Frame::Padding => continue,
                _ => trace_span!("frame", ty = %frame.ty(), path = tracing::field::Empty),
            };

            self.path_stats.get_mut(path_id).frame_rx.record(frame.ty());
            // Crypto, Stream and Datagram frames are special cased in order no pollute
            // the log with payload data
            match &frame {
                Frame::Crypto(f) => {
                    trace!(offset = f.offset, len = f.data.len(), "got frame CRYPTO");
                }
                Frame::Stream(f) => {
                    trace!(id = %f.id, offset = f.offset, len = f.data.len(), fin = f.fin, "got frame STREAM");
                }
                Frame::Datagram(f) => {
                    trace!(len = f.data.len(), "got frame DATAGRAM");
                }
                f => {
                    trace!("got frame {f}");
                }
            }

            let _guard = span.enter();
            if packet.header.is_0rtt() {
                match frame {
                    Frame::Crypto(_) | Frame::Close(Close::Application(_)) => {
                        return Err(TransportError::PROTOCOL_VIOLATION(
                            "illegal frame type in 0-RTT",
                        ));
                    }
                    _ => {
                        if frame.is_1rtt() {
                            return Err(TransportError::PROTOCOL_VIOLATION(
                                "illegal frame type in 0-RTT",
                            ));
                        }
                    }
                }
            }
            ack_eliciting |= frame.is_ack_eliciting();

            // Check whether this could be a probing packet
            match frame {
                Frame::Padding
                | Frame::PathChallenge(_)
                | Frame::PathResponse(_)
                | Frame::NewConnectionId(_)
                | Frame::ObservedAddr(_) => {}
                _ => {
                    is_probing_packet = false;
                }
            }

            match frame {
                Frame::Crypto(frame) => {
                    self.read_crypto(SpaceId::Data, &frame, payload_len)?;
                }
                Frame::Stream(frame) => {
                    if self.streams.received(frame, payload_len)?.should_transmit() {
                        self.spaces[SpaceId::Data].pending.max_data = true;
                    }
                }
                Frame::Ack(ack) => {
                    self.on_ack_received(now, SpaceId::Data, ack)?;
                }
                Frame::PathAck(ack) => {
                    if !self.is_multipath_negotiated() {
                        return Err(TransportError::PROTOCOL_VIOLATION(
                            "received PATH_ACK frame when multipath was not negotiated",
                        ));
                    }
                    span.record("path", tracing::field::display(&ack.path_id));
                    self.on_path_ack_received(now, SpaceId::Data, ack)?;
                }
                Frame::Padding | Frame::Ping => {}
                Frame::Close(reason) => {
                    close = Some(reason);
                }
                Frame::PathChallenge(challenge) => {
                    self.spaces[SpaceKind::Data]
                        .for_path(path_id)
                        .pending_path_responses
                        .push(number, challenge.0, network_path);
                    // If we were passively migrated (e.g. NAT rebinding), our local_ip will
                    // not match. Once we processed a non-probing packet the local_ip will
                    // finally be updated.
                    let path = &mut self
                        .path_mut(path_id)
                        .expect("payload is processed only after the path becomes known");
                    if network_path.remote == path.network_path.remote {
                        // PATH_CHALLENGE on active path, possible off-path packet
                        // forwarding attack. Send a non-probing packet to recover the
                        // active path. See
                        // https://www.rfc-editor.org/rfc/rfc9000.html#section-9.3.3-3. In
                        // rare cases NAT probes might also appear on-path and would also
                        // get a non-probing packet as response. There is little harm in
                        // this.
                        match self.peer_supports_ack_frequency() {
                            true => self.immediate_ack(path_id),
                            false => {
                                self.ping_path(path_id).ok();
                            }
                        }
                    }
                }
                Frame::PathResponse(response) => {
                    // First try to see if this is a NAT probe response.
                    if self
                        .n0_nat_traversal
                        .handle_path_response(network_path, response.0)
                    {
                        self.open_nat_traversed_paths(now);
                    } else {
                        // Try to see if this is a response to an on-path PATH_CHALLENGE.
                        self.handle_path_response_on_path(now, response, path_id);
                    }
                }
                Frame::MaxData(frame::MaxData(bytes)) => {
                    self.streams.received_max_data(bytes);
                }
                Frame::MaxStreamData(frame::MaxStreamData { id, offset }) => {
                    self.streams.received_max_stream_data(id, offset)?;
                }
                Frame::MaxStreams(frame::MaxStreams { dir, count }) => {
                    self.streams.received_max_streams(dir, count)?;
                }
                Frame::ResetStream(frame) => {
                    if self.streams.received_reset(frame)?.should_transmit() {
                        self.spaces[SpaceId::Data].pending.max_data = true;
                    }
                }
                Frame::DataBlocked(DataBlocked(offset)) => {
                    debug!(offset, "peer claims to be blocked at connection level");
                }
                Frame::StreamDataBlocked(StreamDataBlocked { id, offset }) => {
                    if id.initiator() == self.side.side() && id.dir() == Dir::Uni {
                        debug!("got STREAM_DATA_BLOCKED on send-only {}", id);
                        return Err(TransportError::STREAM_STATE_ERROR(
                            "STREAM_DATA_BLOCKED on send-only stream",
                        ));
                    }
                    debug!(
                        stream = %id,
                        offset, "peer claims to be blocked at stream level"
                    );
                }
                Frame::StreamsBlocked(StreamsBlocked { dir, limit }) => {
                    if limit > MAX_STREAM_COUNT {
                        return Err(TransportError::FRAME_ENCODING_ERROR(
                            "unrepresentable stream limit",
                        ));
                    }
                    debug!(
                        "peer claims to be blocked opening more than {} {} streams",
                        limit, dir
                    );
                }
                Frame::StopSending(frame::StopSending { id, error_code }) => {
                    if id.initiator() != self.side.side() {
                        if id.dir() == Dir::Uni {
                            debug!("got STOP_SENDING on recv-only {}", id);
                            return Err(TransportError::STREAM_STATE_ERROR(
                                "STOP_SENDING on recv-only stream",
                            ));
                        }
                    } else if self.streams.is_local_unopened(id) {
                        return Err(TransportError::STREAM_STATE_ERROR(
                            "STOP_SENDING on unopened stream",
                        ));
                    }
                    self.streams.received_stop_sending(id, error_code);
                }
                Frame::RetireConnectionId(frame::RetireConnectionId { path_id, sequence }) => {
                    if let Some(ref path_id) = path_id {
                        span.record("path", tracing::field::display(&path_id));
                    }
                    let path_id = path_id.unwrap_or_default();
                    match self.local_cid_state.get_mut(&path_id) {
                        None => debug!(?path_id, "RETIRE_CONNECTION_ID for unknown path"),
                        Some(cid_state) => {
                            let allow_more_cids = cid_state
                                .on_cid_retirement(sequence, self.peer_params.issue_cids_limit())?;

                            // If the path has closed, we do not issue more CIDs for this path
                            // For details see  https://www.ietf.org/archive/id/draft-ietf-quic-multipath-17.html#section-3.2.2
                            // > an endpoint SHOULD provide new connection IDs for that path, if still open, using PATH_NEW_CONNECTION_ID frames.
                            let has_path = !self.abandoned_paths.contains(&path_id);
                            let allow_more_cids = allow_more_cids && has_path;

                            debug_assert!(!self.state.is_drained()); // required for adding endpoint events, process_payload is never called for drained connections
                            self.endpoint_events
                                .push_back(EndpointEventInner::RetireConnectionId(
                                    now,
                                    path_id,
                                    sequence,
                                    allow_more_cids,
                                ));
                        }
                    }
                }
                Frame::NewConnectionId(frame) => {
                    let path_id = if let Some(path_id) = frame.path_id {
                        if !self.is_multipath_negotiated() {
                            return Err(TransportError::PROTOCOL_VIOLATION(
                                "received PATH_NEW_CONNECTION_ID frame when multipath was not negotiated",
                            ));
                        }
                        if path_id > self.local_max_path_id {
                            return Err(TransportError::PROTOCOL_VIOLATION(
                                "PATH_NEW_CONNECTION_ID contains path_id exceeding current max",
                            ));
                        }
                        path_id
                    } else {
                        PathId::ZERO
                    };

                    if let Some(ref path_id) = frame.path_id {
                        span.record("path", tracing::field::display(&path_id));
                    }

                    if self.abandoned_paths.contains(&path_id) {
                        trace!("ignoring issued CID for abandoned path");
                        continue;
                    }
                    let remote_cids = self
                        .remote_cids
                        .entry(path_id)
                        .or_insert_with(|| CidQueue::new(frame.id));
                    if remote_cids.active().is_empty() {
                        return Err(TransportError::PROTOCOL_VIOLATION(
                            "NEW_CONNECTION_ID when CIDs aren't in use",
                        ));
                    }
                    if frame.retire_prior_to > frame.sequence {
                        return Err(TransportError::PROTOCOL_VIOLATION(
                            "NEW_CONNECTION_ID retiring unissued CIDs",
                        ));
                    }

                    use crate::cid_queue::InsertError;
                    match remote_cids.insert(frame) {
                        Ok(None) => {
                            self.open_nat_traversed_paths(now);
                        }
                        Ok(Some((retired, reset_token))) => {
                            let pending_retired =
                                &mut self.spaces[SpaceId::Data].pending.retire_cids;
                            /// Ensure `pending_retired` cannot grow without bound. Limit is
                            /// somewhat arbitrary but very permissive.
                            const MAX_PENDING_RETIRED_CIDS: u64 = CidQueue::LEN as u64 * 10;
                            // We don't bother counting in-flight frames because those are bounded
                            // by congestion control.
                            if (pending_retired.len() as u64)
                                .saturating_add(retired.end.saturating_sub(retired.start))
                                > MAX_PENDING_RETIRED_CIDS
                            {
                                return Err(TransportError::CONNECTION_ID_LIMIT_ERROR(
                                    "queued too many retired CIDs",
                                ));
                            }
                            pending_retired.extend(retired.map(|seq| (path_id, seq)));
                            self.set_reset_token(path_id, network_path.remote, reset_token);
                            self.open_nat_traversed_paths(now);
                        }
                        Err(InsertError::ExceedsLimit) => {
                            return Err(TransportError::CONNECTION_ID_LIMIT_ERROR(""));
                        }
                        Err(InsertError::Retired) => {
                            trace!("discarding already-retired");
                            // RETIRE_CONNECTION_ID might not have been previously sent if e.g. a
                            // range of connection IDs larger than the active connection ID limit
                            // was retired all at once via retire_prior_to.
                            self.spaces[SpaceId::Data]
                                .pending
                                .retire_cids
                                .push((path_id, frame.sequence));
                            continue;
                        }
                    };

                    if self.side.is_server()
                        && path_id == PathId::ZERO
                        && self
                            .remote_cids
                            .get(&PathId::ZERO)
                            .map(|cids| cids.active_seq() == 0)
                            .unwrap_or_default()
                    {
                        // We're a server still using the initial remote CID for the client, so
                        // let's switch immediately to enable clientside stateless resets.
                        self.update_remote_cid(PathId::ZERO);
                    }
                }
                Frame::NewToken(NewToken { token }) => {
                    let ConnectionSide::Client {
                        token_store,
                        server_name,
                        ..
                    } = &self.side
                    else {
                        return Err(TransportError::PROTOCOL_VIOLATION("client sent NEW_TOKEN"));
                    };
                    if token.is_empty() {
                        return Err(TransportError::FRAME_ENCODING_ERROR("empty token"));
                    }
                    trace!("got new token");
                    token_store.insert(server_name, token);
                }
                Frame::Datagram(datagram) => {
                    if self
                        .datagrams
                        .received(datagram, &self.config.datagram_receive_buffer_size)?
                    {
                        self.events.push_back(Event::DatagramReceived);
                    }
                }
                Frame::AckFrequency(ack_frequency) => {
                    // This frame can only be sent in the Data space

                    if !self.ack_frequency.ack_frequency_received(&ack_frequency)? {
                        // The AckFrequency frame is stale (we have already received a more
                        // recent one)
                        continue;
                    }

                    // Update the params for all of our paths
                    for (path_id, space) in self.spaces[SpaceId::Data].number_spaces.iter_mut() {
                        space.pending_acks.set_ack_frequency_params(&ack_frequency);

                        // Our `max_ack_delay` has been updated, so we may need to adjust
                        // its associated timeout.
                        // Packets received on abandoned paths are always acknowledged immediately.
                        if !self.abandoned_paths.contains(path_id)
                            && let Some(timeout) = space
                                .pending_acks
                                .max_ack_delay_timeout(self.ack_frequency.max_ack_delay)
                        {
                            self.timers.set(
                                Timer::PerPath(*path_id, PathTimer::MaxAckDelay),
                                timeout,
                                self.qlog.with_time(now),
                            );
                        }
                    }
                }
                Frame::ImmediateAck => {
                    // This frame can only be sent in the Data space
                    for pns in self.spaces[SpaceId::Data].iter_paths_mut() {
                        pns.pending_acks.set_immediate_ack_required();
                    }
                }
                Frame::HandshakeDone => {
                    if self.side.is_server() {
                        return Err(TransportError::PROTOCOL_VIOLATION(
                            "client sent HANDSHAKE_DONE",
                        ));
                    }
                    if self.crypto_state.has_keys(EncryptionLevel::Handshake) {
                        self.discard_space(now, SpaceKind::Handshake);
                        self.events.push_back(Event::HandshakeConfirmed);
                        trace!("handshake confirmed");
                    }
                }
                Frame::ObservedAddr(observed) => {
                    // check if params allows the peer to send report and this node to receive it
                    trace!(seq_no = %observed.seq_no, ip = %observed.ip, port = observed.port);
                    if !self
                        .peer_params
                        .address_discovery_role
                        .should_report(&self.config.address_discovery_role)
                    {
                        return Err(TransportError::PROTOCOL_VIOLATION(
                            "received OBSERVED_ADDRESS frame when not negotiated",
                        ));
                    }
                    // must only be sent in data space
                    if packet.header.space() != SpaceKind::Data {
                        return Err(TransportError::PROTOCOL_VIOLATION(
                            "OBSERVED_ADDRESS frame outside data space",
                        ));
                    }

                    let space_open_status =
                        self.spaces[SpaceKind::Data].for_path(path_id).open_status;
                    let path = self.path_data_mut(path_id);
                    if path.network_path.remote == network_path.remote {
                        if let Some(updated) = path.update_observed_addr_report(observed)
                            && space_open_status == OpenStatus::Informed
                        {
                            self.events.push_back(Event::Path(PathEvent::ObservedAddr {
                                id: path_id,
                                addr: updated,
                            }));
                            // otherwise the event is reported when the path is deemed open
                        }
                    } else {
                        // include in migration
                        migration_observed_addr = Some(observed)
                    }
                }
                Frame::PathAbandon(frame::PathAbandon {
                    path_id,
                    error_code,
                }) => {
                    span.record("path", tracing::field::display(&path_id));
                    match self.close_path_inner(
                        now,
                        path_id,
                        PathAbandonReason::RemoteAbandoned {
                            error_code: error_code.into(),
                        },
                    ) {
                        Ok(()) => {
                            trace!("peer abandoned path");
                        }
                        Err(ClosePathError::ClosedPath) => {
                            trace!("peer abandoned already closed path");
                        }
                        Err(ClosePathError::MultipathNotNegotiated) => {
                            return Err(TransportError::PROTOCOL_VIOLATION(
                                "received PATH_ABANDON frame when multipath was not negotiated",
                            ));
                        }
                        Err(ClosePathError::LastOpenPath) => {
                            // Not reachable: close_path_inner allows remote abandons
                            // for the last path. But handle gracefully just in case.
                            error!(
                                "peer abandoned last path but close_path_inner returned LastOpenPath"
                            );
                        }
                    };

                    // Start draining the path if it still exists and hasn't started draining yet.
                    if let Some(path) = self.paths.get_mut(&path_id)
                        && !mem::replace(&mut path.data.draining, true)
                    {
                        let ack_delay = self.ack_frequency.max_ack_delay_for_pto();
                        let pto = path.data.rtt.pto_base() + ack_delay;
                        self.timers.set(
                            Timer::PerPath(path_id, PathTimer::PathDrained),
                            now + 3 * pto,
                            self.qlog.with_time(now),
                        );

                        self.set_max_path_id(now, self.local_max_path_id.saturating_add(1u8));
                    }
                }
                Frame::PathStatusAvailable(info) => {
                    span.record("path", tracing::field::display(&info.path_id));
                    if self.is_multipath_negotiated() {
                        self.on_path_status(
                            info.path_id,
                            PathStatus::Available,
                            info.status_seq_no,
                        );
                    } else {
                        return Err(TransportError::PROTOCOL_VIOLATION(
                            "received PATH_STATUS_AVAILABLE frame when multipath was not negotiated",
                        ));
                    }
                }
                Frame::PathStatusBackup(info) => {
                    span.record("path", tracing::field::display(&info.path_id));
                    if self.is_multipath_negotiated() {
                        self.on_path_status(info.path_id, PathStatus::Backup, info.status_seq_no);
                    } else {
                        return Err(TransportError::PROTOCOL_VIOLATION(
                            "received PATH_STATUS_BACKUP frame when multipath was not negotiated",
                        ));
                    }
                }
                Frame::MaxPathId(frame::MaxPathId(path_id)) => {
                    span.record("path", tracing::field::display(&path_id));
                    if !self.is_multipath_negotiated() {
                        return Err(TransportError::PROTOCOL_VIOLATION(
                            "received MAX_PATH_ID frame when multipath was not negotiated",
                        ));
                    }
                    // frames that do not increase the path id are ignored
                    if path_id > self.remote_max_path_id {
                        self.remote_max_path_id = path_id;
                        self.issue_first_path_cids(now);
                        self.open_nat_traversed_paths(now);
                    }
                }
                Frame::PathsBlocked(frame::PathsBlocked(max_path_id)) => {
                    // Receipt of a value of Maximum Path Identifier or Path Identifier that is higher than the local maximum value MUST
                    // be treated as a connection error of type PROTOCOL_VIOLATION.
                    // Ref <https://www.ietf.org/archive/id/draft-ietf-quic-multipath-14.html#name-paths_blocked-and-path_cids>
                    if self.is_multipath_negotiated() {
                        if max_path_id > self.local_max_path_id {
                            return Err(TransportError::PROTOCOL_VIOLATION(
                                "PATHS_BLOCKED maximum path identifier was larger than local maximum",
                            ));
                        }
                    } else {
                        return Err(TransportError::PROTOCOL_VIOLATION(
                            "received PATHS_BLOCKED frame when not multipath was not negotiated",
                        ));
                    }
                }
                Frame::PathCidsBlocked(frame::PathCidsBlocked { path_id, next_seq }) => {
                    // Nothing to do.  This is recorded in the frame stats, but otherwise we
                    // always issue all CIDs we're allowed to issue, so either this is an
                    // impatient peer or a bug on our side.

                    // Receipt of a value of Maximum Path Identifier or Path Identifier that is higher than the local maximum value MUST
                    // be treated as a connection error of type PROTOCOL_VIOLATION.
                    // Ref <https://www.ietf.org/archive/id/draft-ietf-quic-multipath-14.html#name-paths_blocked-and-path_cids>
                    if self.is_multipath_negotiated() {
                        if path_id > self.local_max_path_id {
                            return Err(TransportError::PROTOCOL_VIOLATION(
                                "PATH_CIDS_BLOCKED path identifier was larger than local maximum",
                            ));
                        }
                        if self
                            .local_cid_state
                            .get(&path_id)
                            // The PATH_CIDS_BLOCKED frame may arrive after we've discarded the path state.
                            // In that case, we can't check for the protocol violation.
                            .is_some_and(|cid_state| next_seq.0 > cid_state.active_seq().1 + 1)
                        {
                            return Err(TransportError::PROTOCOL_VIOLATION(
                                "PATH_CIDS_BLOCKED next sequence number larger than in local state",
                            ));
                        }
                        debug!(%path_id, %next_seq, "received PATH_CIDS_BLOCKED");
                    } else {
                        return Err(TransportError::PROTOCOL_VIOLATION(
                            "received PATH_CIDS_BLOCKED frame when not multipath was not negotiated",
                        ));
                    }
                }
                Frame::AddAddress(addr) => {
                    let client_state = match self.n0_nat_traversal.client_side_mut() {
                        Ok(state) => state,
                        Err(err) => {
                            return Err(TransportError::PROTOCOL_VIOLATION(format!(
                                "Nat traversal(ADD_ADDRESS): {err}"
                            )));
                        }
                    };

                    if !client_state.check_remote_address(&addr) {
                        // if the address is not valid we flag it, but update anyway
                        warn!(?addr, "server sent illegal ADD_ADDRESS frame");
                    }

                    match client_state.add_remote_address(addr) {
                        Ok(maybe_added) => {
                            if let Some(added) = maybe_added {
                                self.events.push_back(Event::NatTraversal(
                                    n0_nat_traversal::Event::AddressAdded(added),
                                ));
                            }
                        }
                        Err(e) => {
                            warn!(%e, "failed to add remote address")
                        }
                    }
                }
                Frame::RemoveAddress(addr) => {
                    let client_state = match self.n0_nat_traversal.client_side_mut() {
                        Ok(state) => state,
                        Err(err) => {
                            return Err(TransportError::PROTOCOL_VIOLATION(format!(
                                "Nat traversal(REMOVE_ADDRESS): {err}"
                            )));
                        }
                    };
                    if let Some(removed_addr) = client_state.remove_remote_address(addr) {
                        self.events.push_back(Event::NatTraversal(
                            n0_nat_traversal::Event::AddressRemoved(removed_addr),
                        ));
                    }
                }
                Frame::ReachOut(reach_out) => {
                    let ipv6 = self.is_ipv6();
                    let server_state = match self.n0_nat_traversal.server_side_mut() {
                        Ok(state) => state,
                        Err(err) => {
                            return Err(TransportError::PROTOCOL_VIOLATION(format!(
                                "Nat traversal(REACH_OUT): {err}"
                            )));
                        }
                    };

                    let round_before = server_state.current_round();

                    if let Err(err) = server_state.handle_reach_out(reach_out, ipv6) {
                        return Err(TransportError::PROTOCOL_VIOLATION(format!(
                            "Nat traversal(REACH_OUT): {err}"
                        )));
                    }

                    if server_state.current_round() > round_before {
                        // A new round was started, reset the NAT probe retry timer.
                        if let Some(delay) =
                            self.n0_nat_traversal.retry_delay(self.config.initial_rtt)
                        {
                            self.timers.set(
                                Timer::Conn(ConnTimer::NatTraversalProbeRetry),
                                now + delay,
                                self.qlog.with_time(now),
                            );
                        }
                    }
                }
            }
        }

        let space = self.spaces[SpaceId::Data].for_path(path_id);
        if space
            .pending_acks
            .packet_received(now, number, ack_eliciting, &space.dedup)
        {
            if self.abandoned_paths.contains(&path_id) {
                // § 3.4.3 QUIC-MULTIPATH: promptly send ACKs for packets received from
                // abandoned paths.
                space.pending_acks.set_immediate_ack_required();
            } else {
                self.timers.set(
                    Timer::PerPath(path_id, PathTimer::MaxAckDelay),
                    now + self.ack_frequency.max_ack_delay,
                    self.qlog.with_time(now),
                );
            }
        }

        // Issue stream ID credit due to ACKs of outgoing finish/resets and incoming finish/resets
        // on stopped streams. Incoming finishes/resets on open streams are not handled here as they
        // are only freed, and hence only issue credit, once the application has been notified
        // during a read on the stream.
        let pending = &mut self.spaces[SpaceId::Data].pending;
        self.streams.queue_max_stream_id(pending);

        if let Some(reason) = close {
            self.state
                .move_to_draining(Some(reason.into()), &mut self.endpoint_events);
            self.connection_close_pending = true;
        }

        // For Multipath any packet triggers migration. For RFC9000 or QNT (+ Multipath)
        // only non-probing packets trigger migration.
        let migrate_on_any_packet =
            self.is_multipath_negotiated() && !self.n0_nat_traversal.is_negotiated();

        // Only migrate if this is the largest packet number seen.
        let is_largest_received_pn = Some(number)
            == self.spaces[SpaceId::Data]
                .for_path(path_id)
                .largest_received_packet_number;

        // If we receive a non-probing packet on a new local IP that means we had a NAT
        // rebinding-like migration. We update our local address but do not otherwise
        // validate the new path, we only need to validate the path if the peer migrates per
        // RFC9000 §9: https://www.rfc-editor.org/rfc/rfc9000.html#section-9-4
        if (migrate_on_any_packet || !is_probing_packet)
            && is_largest_received_pn
            && self.local_ip_may_migrate()
            && let Some(new_local_ip) = network_path.local_ip
        {
            let path_data = self.path_data_mut(path_id);
            if path_data
                .network_path
                .local_ip
                .is_some_and(|ip| ip != new_local_ip)
            {
                debug!(
                    %path_id,
                    new_4tuple = %network_path,
                    prev_4tuple = %path_data.network_path,
                    "local address passive migration"
                );
            }
            path_data.network_path.local_ip = Some(new_local_ip)
        }

        // If the peer migrated to a new address, trigger migration.
        if self.peer_may_migrate()
            && (migrate_on_any_packet || !is_probing_packet)
            && is_largest_received_pn
            && network_path.remote != self.path_data(path_id).network_path.remote
        {
            self.migrate(path_id, now, network_path, migration_observed_addr);
            // Break linkability, if possible
            self.update_remote_cid(path_id);
            self.spin = false;
        }

        Ok(())
    }

    /// Handles any on-path PATH_RESPONSE frames.
    ///
    /// *path_id* and *network_path* are those on which the PATH_RESPONSE was received.
    fn handle_path_response_on_path(
        &mut self,
        now: Instant,
        response: frame::PathResponse,
        path_id: PathId,
    ) {
        let is_multipath_negotiated = self.is_multipath_negotiated();
        let path = self
            .paths
            .get_mut(&path_id)
            .expect("payload is processed only after the path becomes known");
        match path.data.on_path_response_received(now, response.0) {
            paths::OnPathResponseReceived::OnPath if !self.abandoned_paths.contains(&path_id) => {
                let qlog = self.qlog.with_time(now);
                self.timers.stop(
                    Timer::PerPath(path_id, PathTimer::PathValidationFailed),
                    qlog.clone(),
                );
                let next_challenge = path
                    .data
                    .earliest_on_path_expiring_challenge()
                    .map(|time| time + self.ack_frequency.max_ack_delay_for_pto());
                self.timers.set_or_stop(
                    Timer::PerPath(path_id, PathTimer::PathChallengeLost),
                    next_challenge,
                    qlog,
                );
                let pns = self.spaces[SpaceKind::Data].for_path(path_id);
                if !matches!(pns.open_status, OpenStatus::Informed) {
                    if is_multipath_negotiated {
                        self.events
                            .push_back(Event::Path(PathEvent::Established { id: path_id }));
                    }
                    pns.open_status = OpenStatus::Informed;
                    if let Some(observed) = path.data.last_observed_addr_report.as_ref() {
                        self.events.push_back(Event::Path(PathEvent::ObservedAddr {
                            id: path_id,
                            addr: observed.socket_addr(),
                        }));
                    }
                }
                if let Some((_, ref mut prev)) = path.prev {
                    // If an on-path response was received while there is a
                    // previous path from a migration, then the new path is
                    // validated and we can stop sending challenges that try to
                    // re-validate the previous path.
                    prev.reset_on_path_challenges();
                }
            }
            paths::OnPathResponseReceived::OnPath => {
                trace!(
                    %response,
                    "ignoring PATH_RESPONSE received after path is abandoned"
                );
            }
            paths::OnPathResponseReceived::Unknown => {
                debug!(%response, "ignoring invalid PATH_RESPONSE");
            }
            paths::OnPathResponseReceived::Ignored {
                sent_on,
                current_path,
            } => {
                debug!(%sent_on, %current_path, %response, "ignoring valid PATH_RESPONSE");
            }
        }
    }

    /// Opens any paths that have been successfully NAT traversed.
    fn open_nat_traversed_paths(&mut self, now: Instant) {
        while let Some(network_path) = self
            .n0_nat_traversal
            .client_side_mut()
            .ok()
            .and_then(|s| s.pop_pending_path_open())
        {
            match self.open_path_ensure(network_path, PathStatus::Backup, now) {
                Ok((path_id, already_existed)) => {
                    debug!(
                        %path_id,
                        ?network_path,
                        new_path = !already_existed,
                        "Opened NAT traversal path",
                    );
                }
                Err(err) => match err {
                    PathError::MultipathNotNegotiated
                    | PathError::ServerSideNotAllowed
                    | PathError::ValidationFailed
                    | PathError::InvalidRemoteAddress(_) => {
                        error!(
                            ?err,
                            ?network_path,
                            "Failed to open path for successful NAT traversal"
                        );
                    }
                    PathError::MaxPathIdReached | PathError::RemoteCidsExhausted => {
                        // Temporary error, put back.
                        self.n0_nat_traversal
                            .client_side_mut()
                            .map(|s| s.push_pending_path_open(network_path))
                            .ok();
                        debug!(
                            ?err,
                            ?network_path,
                            "Blocked opening NAT traversal path, enqueued"
                        );
                        return;
                    }
                },
            }
        }
    }

    /// Migrates the 4-tuple of the path.
    ///
    /// This creates a new [`PathData`] for the migrated path and stores the previous
    /// [`PathData`] in [`PathState::prev`].
    fn migrate(
        &mut self,
        path_id: PathId,
        now: Instant,
        network_path: FourTuple,
        observed_addr: Option<ObservedAddr>,
    ) {
        trace!(
            new_4tuple = %network_path,
            prev_4tuple = %self.path_data(path_id).network_path,
            %path_id,
            "migration initiated",
        );
        self.path_generation_counter = self.path_generation_counter.wrapping_add(1);
        // TODO(@divma): conditions for path migration in multipath are very specific, check them
        // again to prevent path migrations that should actually create a new path

        // Reset rtt/congestion state for new path unless it looks like a NAT rebinding.
        // Note that the congestion window will not grow until validation terminates. Helps mitigate
        // amplification attacks performed by spoofing source addresses.
        let prev_pto = self.pto(SpaceKind::Data, path_id);
        let path = self.paths.get_mut(&path_id).expect("known path");
        let mut new_path_data = if network_path.remote.is_ipv4()
            && network_path.remote.ip() == path.data.network_path.remote.ip()
        {
            PathData::from_previous(network_path, &path.data, self.path_generation_counter, now)
        } else {
            let peer_max_udp_payload_size =
                u16::try_from(self.peer_params.max_udp_payload_size.into_inner())
                    .unwrap_or(u16::MAX);
            PathData::new(
                network_path,
                self.allow_mtud,
                Some(peer_max_udp_payload_size),
                self.path_generation_counter,
                now,
                &self.config,
            )
        };
        new_path_data.last_observed_addr_report = path.data.last_observed_addr_report.clone();
        if let Some(report) = observed_addr
            && let Some(updated) = new_path_data.update_observed_addr_report(report)
        {
            tracing::info!("adding observed addr event from migration");
            self.events.push_back(Event::Path(PathEvent::ObservedAddr {
                id: path_id,
                addr: updated,
            }));
        }
        new_path_data.pending_challenge = true;
        new_path_data.pending.observed_address = self
            .config
            .address_discovery_role
            .should_report(&self.peer_params.address_discovery_role);

        let mut prev_path_data = mem::replace(&mut path.data, new_path_data);

        // Only store this as previous path if it was validated. For all we know there could
        // already be a previous path stored which might have been validated in the past,
        // which is more valuable than one that's not yet validated.
        //
        // With multipath it is possible that there are no remote CIDs for the path ID
        // yet. In this case we would never have sent on this path yet and would not be able
        // to send a PATH_CHALLENGE either, which is currently a fire-and-forget affair
        // anyway. So don't store such a path either.
        if !prev_path_data.validated
            && let Some(cid) = self.remote_cids.get(&path_id).map(CidQueue::active)
        {
            prev_path_data.pending_challenge = true;
            // We haven't updated the remote CID yet, this captures the remote CID we were using on
            // the previous path.
            path.prev = Some((cid, prev_path_data));
        }

        // We need to re-assign the correct remote to this path in qlog
        self.qlog.emit_tuple_assigned(path_id, network_path, now);

        self.timers.set(
            Timer::PerPath(path_id, PathTimer::PathValidationFailed),
            now + 3 * cmp::max(self.pto(SpaceKind::Data, path_id), prev_pto),
            self.qlog.with_time(now),
        );
    }

    /// Handle a change in the local address, i.e. an active migration
    ///
    /// In the general (non-multipath) case, paths will perform a RFC9000 migration and be pinged
    /// for a liveness check. This is the behaviour of a path assumed to be recoverable, even if
    /// this is not the case.
    ///
    /// Clients in a connection in which multipath has been negotiated should migrate paths to new
    /// [`PathId`]s. For paths that are known to be non-recoverable can be migrated to a new
    /// [`PathId`] by closing the current path, and opening a new one to the same remote. Treating
    /// paths as non recoverable when necessary accelerates connectivity re-establishment, or might
    /// allow it altogether.
    ///
    /// The optional `hint` allows callers to indicate when paths are non-recoverable and should be
    /// migrated to new a [`PathId`].
    // NOTE: only clients are allowed to migrate, but generally dealing with RFC9000 migrations is
    // lacking <https://github.com/n0-computer/noq/issues/364>
    pub fn handle_network_change(&mut self, hint: Option<&dyn NetworkChangeHint>, now: Instant) {
        debug!("network changed");
        if self.state.is_drained() {
            return;
        }
        if self.highest_space < SpaceKind::Data {
            for path in self.paths.values_mut() {
                // Clear the local address for it to be obtained from the socket again.
                path.data.network_path.local_ip = None;
            }

            self.update_remote_cid(PathId::ZERO);
            self.ping();

            return;
        }

        // Paths that can't recover so a new path should be open instead. If multipath is not
        // negotiated, this will be empty.
        let mut non_recoverable_paths = Vec::default();
        let mut recoverable_paths = Vec::default();
        let mut open_paths = 0;

        let is_multipath_negotiated = self.is_multipath_negotiated();
        let is_client = self.side().is_client();
        let immediate_ack_allowed = self.peer_supports_ack_frequency();

        for (path_id, path) in self.paths.iter_mut() {
            if self.abandoned_paths.contains(path_id) {
                continue;
            }
            open_paths += 1;

            // Read the network path BEFORE clearing local_ip, so the hint can
            // check which interface the path was using.
            let network_path = path.data.network_path;

            // Clear the local address for it to be obtained from the socket again. This applies to
            // all paths, regardless of being considered recoverable or not
            path.data.network_path.local_ip = None;
            let remote = network_path.remote;

            // Without multipath, the connection tries to recover the single path, whereas with
            // multipath, even in a single-path scenario, we attempt to migrate the path to a new
            // PathId.
            let attempt_to_recover = if is_multipath_negotiated {
                // Use the hint to determine if the path can recover. When no hint is
                // provided, clients default to non-recoverable (abandon and re-open)
                // while servers default to recoverable (attempt in-place recovery).
                hint.map(|h| h.is_path_recoverable(*path_id, network_path))
                    .unwrap_or(!is_client)
            } else {
                // In the non multipath case, we try to recover the single active path
                true
            };

            if attempt_to_recover {
                recoverable_paths.push((*path_id, remote));
            } else {
                non_recoverable_paths.push((*path_id, remote, path.data.local_status()))
            }
        }

        /* NON RECOVERABLE PATHS */
        // This are handled first, so that in case the treatment intended for these fails, we can
        // go the recoverable route instead.

        // Decide if we need to close first or open first in the multipath case.
        // - Opening first has a higher risk of getting limited by the negotiated MAX_PATH_ID.
        // - Closing first risks this being the only open path.
        // We prefer closing paths first unless we identify this is the last open path.
        let open_first = open_paths == non_recoverable_paths.len();

        for (path_id, remote, status) in non_recoverable_paths.into_iter() {
            let network_path = FourTuple {
                remote,
                local_ip: None, /* allow the local ip to be discovered */
            };

            if open_first && let Err(e) = self.open_path(network_path, status, now) {
                if self.side().is_client() {
                    debug!(%e, "Failed to open new path for network change");
                }
                // if this fails, let the path try to recover itself
                recoverable_paths.push((path_id, remote));
                continue;
            }

            if let Err(e) =
                self.close_path_inner(now, path_id, PathAbandonReason::UnusableAfterNetworkChange)
            {
                debug!(%e,"Failed to close unrecoverable path after network change");
                recoverable_paths.push((path_id, remote));
                continue;
            }

            if !open_first && let Err(e) = self.open_path(network_path, status, now) {
                // Path has already been closed if we got here. Since the path was not recoverable,
                // this might be desirable in any case, because other paths exist (!open_first) and
                // this was is considered non recoverable
                debug!(%e,"Failed to open new path for network change");
            }
        }

        /* RECOVERABLE PATHS */

        for (path_id, remote) in recoverable_paths.into_iter() {
            // Schedule a Ping for a liveness check.
            if let Some(path_space) = self.spaces[SpaceId::Data].number_spaces.get_mut(&path_id) {
                path_space.pending_ping = true;

                if immediate_ack_allowed {
                    path_space.pending_immediate_ack = true;
                }
            }

            // Reset PTO backoff so retransmits resume promptly. Congestion controller and
            // RTT are intentionally preserved for recoverable paths. We explicitly allow
            // this reset also during the handshake, so do not check
            // Self::peer_competed_handshake_address_validation.
            if let Some(path) = self.paths.get_mut(&path_id) {
                path.data.pto_count = 0;
            }
            self.set_loss_detection_timer(now, path_id);

            let Some((reset_token, retired)) =
                self.remote_cids.get_mut(&path_id).and_then(CidQueue::next)
            else {
                continue;
            };

            // Retire the current remote CID and any CIDs we had to skip.
            self.spaces[SpaceId::Data]
                .pending
                .retire_cids
                .extend(retired.map(|seq| (path_id, seq)));

            debug_assert!(!self.state.is_drained()); // required for endpoint_events, checked above
            self.endpoint_events
                .push_back(EndpointEventInner::ResetToken(path_id, remote, reset_token));
        }
    }

    /// Switch to a previously unused remote connection ID, if possible
    fn update_remote_cid(&mut self, path_id: PathId) {
        let Some((reset_token, retired)) = self
            .remote_cids
            .get_mut(&path_id)
            .and_then(|cids| cids.next())
        else {
            return;
        };

        // Retire the current remote CID and any CIDs we had to skip.
        self.spaces[SpaceId::Data]
            .pending
            .retire_cids
            .extend(retired.map(|seq| (path_id, seq)));
        let remote = self.path_data(path_id).network_path.remote;
        self.set_reset_token(path_id, remote, reset_token);
    }

    /// Sends this reset token to the endpoint
    ///
    /// The endpoint needs to know the reset tokens issued by the peer, so that if the peer
    /// sends a reset token it knows to route it to this connection. See RFC 9000 section
    /// 10.3. Stateless Reset.
    ///
    /// Reset tokens are different for each path, the endpoint identifies paths by peer
    /// socket address however, not by path ID.
    fn set_reset_token(&mut self, path_id: PathId, remote: SocketAddr, reset_token: ResetToken) {
        debug_assert!(!self.state.is_drained()); // required for endpoint events, set_reset_token is never called for drained connections
        self.endpoint_events
            .push_back(EndpointEventInner::ResetToken(path_id, remote, reset_token));

        // During the handshake the server sends a reset token in the transport
        // parameters. When we are the client and we receive the reset token during the
        // handshake we want this to affect our peer transport parameters.
        // TODO(flub): Pretty sure this is pointless, the entire params is overwritten
        //    shortly after this was called.  And then the params don't have this anymore.
        if path_id == PathId::ZERO {
            self.peer_params.stateless_reset_token = Some(reset_token);
        }
    }

    /// Issue an initial set of connection IDs to the peer upon connection
    fn issue_first_cids(&mut self, now: Instant) {
        if self
            .local_cid_state
            .get(&PathId::ZERO)
            .expect("PathId::ZERO exists when the connection is created")
            .cid_len()
            == 0
        {
            return;
        }

        // Subtract 1 to account for the CID we supplied while handshaking
        let mut n = self.peer_params.issue_cids_limit() - 1;
        if let ConnectionSide::Server { server_config } = &self.side
            && server_config.has_preferred_address()
        {
            // We also sent a CID in the transport parameters
            n -= 1;
        }
        debug_assert!(!self.state.is_drained()); // requirement for endpoint_events
        self.endpoint_events
            .push_back(EndpointEventInner::NeedIdentifiers(PathId::ZERO, now, n));
    }

    /// Issues an initial set of CIDs for paths that have not yet had any CIDs issued
    ///
    /// Later CIDs are issued when CIDs expire or are retired by the peer.
    fn issue_first_path_cids(&mut self, now: Instant) {
        if let Some(max_path_id) = self.max_path_id() {
            let mut path_id = self.max_path_id_with_cids.next();
            while path_id <= max_path_id {
                self.endpoint_events
                    .push_back(EndpointEventInner::NeedIdentifiers(
                        path_id,
                        now,
                        self.peer_params.issue_cids_limit(),
                    ));
                path_id = path_id.next();
            }
            self.max_path_id_with_cids = max_path_id;
        }
    }

    /// Populates a packet with frames
    ///
    /// This tries to fit as many frames as possible into the packet.
    ///
    /// *path_exclusive_only* means to only build frames which can only be sent on this
    /// *path.  This is used in multipath for backup paths while there is still an active
    /// *path.
    fn populate_packet<'a, 'b>(
        &mut self,
        now: Instant,
        space_id: SpaceId,
        path_id: PathId,
        scheduling_info: &PathSchedulingInfo,
        builder: &mut PacketBuilder<'a, 'b>,
    ) {
        let is_multipath_negotiated = self.is_multipath_negotiated();
        let space_has_keys = self.crypto_state.has_keys(space_id.encryption_level());
        let is_0rtt = space_id == SpaceId::Data && !space_has_keys;
        let stats = &mut self.path_stats.get_mut(path_id).frame_tx;
        let space = &mut self.spaces[space_id];
        let path = &mut self.paths.get_mut(&path_id).expect("known path").data;
        space
            .for_path(path_id)
            .pending_acks
            .maybe_ack_non_eliciting();

        // HANDSHAKE_DONE
        if !is_0rtt
            && !scheduling_info.is_abandoned
            && scheduling_info.may_send_data
            && mem::replace(&mut space.pending.handshake_done, false)
        {
            builder.write_frame(frame::HandshakeDone, stats);
        }

        // PING
        if !scheduling_info.is_abandoned
            && mem::replace(&mut space.for_path(path_id).pending_ping, false)
        {
            builder.write_frame(frame::Ping, stats);
        }

        // IMMEDIATE_ACK
        if !scheduling_info.is_abandoned
            && mem::replace(&mut space.for_path(path_id).pending_immediate_ack, false)
        {
            debug_assert_eq!(
                space_id,
                SpaceId::Data,
                "immediate acks must be sent in the data space"
            );
            builder.write_frame(frame::ImmediateAck, stats);
        }

        // ACK
        if !scheduling_info.is_abandoned && scheduling_info.may_send_data {
            for path_id in space
                .number_spaces
                .iter_mut()
                .filter(|(_, pns)| pns.pending_acks.can_send())
                .map(|(&path_id, _)| path_id)
                .collect::<Vec<_>>()
            {
                Self::populate_acks(
                    now,
                    self.receiving_ecn,
                    path_id,
                    space_id,
                    space,
                    is_multipath_negotiated,
                    builder,
                    stats,
                    space_has_keys,
                );
            }
        }

        // ACK_FREQUENCY
        if !scheduling_info.is_abandoned
            && scheduling_info.may_send_data
            && mem::replace(&mut space.pending.ack_frequency, false)
        {
            let sequence_number = self.ack_frequency.next_sequence_number();

            // Safe to unwrap because this is always provided when ACK frequency is enabled
            let config = self.config.ack_frequency_config.as_ref().unwrap();

            // Ensure the delay is within bounds to avoid a PROTOCOL_VIOLATION error
            let max_ack_delay = self.ack_frequency.candidate_max_ack_delay(
                path.rtt.get(),
                config,
                &self.peer_params,
            );

            let frame = frame::AckFrequency {
                sequence: sequence_number,
                ack_eliciting_threshold: config.ack_eliciting_threshold,
                request_max_ack_delay: max_ack_delay.as_micros().try_into().unwrap_or(VarInt::MAX),
                reordering_threshold: config.reordering_threshold,
            };
            builder.write_frame(frame, stats);

            self.ack_frequency
                .ack_frequency_sent(path_id, builder.packet_number, max_ack_delay);
        }

        // PATH_CHALLENGE (on-path)
        if !scheduling_info.is_abandoned
            && space_id == SpaceId::Data
            && path.pending_challenge
            // we don't want to send new challenges if we are already closing
            && !self.state.is_closed()
            && builder.frame_space_remaining() > frame::PathChallenge::SIZE_BOUND
            // An on-path PATH_CHALLENGE must be part of datagrams expanded to the
            // MIN_INITIAL_SIZE (1200 bytes).
            && builder.buf.segment_size() >= usize::from(MIN_INITIAL_SIZE)
        {
            path.pending_challenge = false;

            let token = self.rng.random();
            path.record_path_challenge_sent(now, token, path.network_path);
            // Generate a new challenge every time we send a new PATH_CHALLENGE
            let challenge = frame::PathChallenge(token);
            builder.write_frame(challenge, stats);
            builder.require_padding();

            // On-path challenges need a PATH_RESPONSE and not only an ACK which can be
            // received on any path. So set a timer manually instead of relying on the usual
            // LossDetection/PTO timer. This timer interval keeps exponentially increasing
            // with missing responses like the normal PTO interval.
            self.timers.set(
                Timer::PerPath(path_id, PathTimer::PathChallengeLost),
                now + path.on_path_challenge_pto(),
                self.qlog.with_time(now),
            );

            if is_multipath_negotiated && !path.validated && path.pending_challenge {
                // queue informing the path status along with the challenge
                space.pending.path_status.insert(path_id);
            }

            // Always include an OBSERVED_ADDR frame with a PATH_CHALLENGE, regardless
            // of whether one has already been sent on this path.
            path.pending.observed_address = self
                .config
                .address_discovery_role
                .should_report(&self.peer_params.address_discovery_role);
        }

        // PATH_RESPONSE (on-path)
        if !scheduling_info.is_abandoned
            && space_id == SpaceId::Data
            && builder.frame_space_remaining() > frame::PathResponse::SIZE_BOUND
            // An on-path PATH_RESPONSE must be part of datagrams expanded to the
            // MIN_INITIAL_SIZE (1200 bytes).
            && builder.buf.segment_size() >= usize::from(MIN_INITIAL_SIZE)
            && let Some(token) = space.for_path(path_id).pending_path_responses.pop_on_path(path.network_path)
        {
            let response = frame::PathResponse(token);
            builder.write_frame(response, stats);
            builder.require_padding();

            // NOTE: this is technically not required but might be useful to ride the
            // request/response nature of path challenges to refresh an observation
            // Since PATH_RESPONSE is a probing frame, this is allowed by the spec.
            path.pending.observed_address = self
                .config
                .address_discovery_role
                .should_report(&self.peer_params.address_discovery_role);
        }

        // ADD_ADDRESS
        while space_id == SpaceId::Data
            && !scheduling_info.is_abandoned
            && scheduling_info.may_send_data
            && frame::AddAddress::SIZE_BOUND <= builder.frame_space_remaining()
        {
            if let Some(added_address) = space.pending.add_address.pop_last() {
                builder.write_frame(added_address, stats);
            } else {
                break;
            }
        }

        // REMOVE_ADDRESS
        while space_id == SpaceId::Data
            && !scheduling_info.is_abandoned
            && scheduling_info.may_send_data
            && frame::RemoveAddress::SIZE_BOUND <= builder.frame_space_remaining()
        {
            if let Some(removed_address) = space.pending.remove_address.pop_last() {
                builder.write_frame(removed_address, stats);
            } else {
                break;
            }
        }

        // REACH_OUT
        while !scheduling_info.is_abandoned
            && scheduling_info.may_send_data
            && let Some(reach_out) = space
                .pending
                .reach_out
                .pop_if(|frame| builder.frame_space_remaining() >= frame.size())
        {
            builder.write_frame(reach_out, stats);
        }

        // PATH_ABANDON
        if space_id == SpaceId::Data
            && scheduling_info.is_abandoned
            && scheduling_info.may_self_abandon
            && frame::PathAbandon::SIZE_BOUND <= builder.frame_space_remaining()
            && let Some(error_code) = space.pending.path_abandon.remove(&path_id)
        {
            let frame = frame::PathAbandon {
                path_id,
                error_code,
            };
            builder.write_frame(frame, stats);

            // Consider remotely issued CIDs as retired now that we have sent this frame at
            // least once.
            self.remote_cids.remove(&path_id);
        }
        while space_id == SpaceId::Data
            && scheduling_info.may_send_data
            && frame::PathAbandon::SIZE_BOUND <= builder.frame_space_remaining()
            && let Some((abandoned_path_id, error_code)) = space.pending.path_abandon.pop_first()
        {
            let frame = frame::PathAbandon {
                path_id: abandoned_path_id,
                error_code,
            };
            builder.write_frame(frame, stats);

            // Consider remotely issued CIDs as retired now that we have sent this frame at
            // least once.
            self.remote_cids.remove(&abandoned_path_id);
        }

        // OBSERVED_ADDR
        if !scheduling_info.is_abandoned
            && space_id == SpaceId::Data
            && path.pending.observed_address
        {
            let frame = ObservedAddr::new(path.network_path.remote, self.next_observed_addr_seq_no);
            if builder.frame_space_remaining() > frame.size() {
                builder.write_frame(frame, stats);

                self.next_observed_addr_seq_no = self.next_observed_addr_seq_no.saturating_add(1u8);
                path.pending.observed_address = false;
            }
        }

        // CRYPTO
        while !is_0rtt
            && !scheduling_info.is_abandoned
            && scheduling_info.may_send_data
            && builder.frame_space_remaining() > frame::Crypto::SIZE_BOUND
        {
            let Some(mut frame) = space.pending.crypto.pop_front() else {
                break;
            };

            // Calculate the maximum amount of crypto data we can store in the buffer.
            // Since the offset is known, we can reserve the exact size required to encode it.
            // For length we reserve 2bytes which allows to encode up to 2^14,
            // which is more than what fits into normally sized QUIC frames.
            let max_crypto_data_size = builder.frame_space_remaining()
                - 1 // Frame Type
                - VarInt::size(unsafe { VarInt::from_u64_unchecked(frame.offset) })
                - 2; // Maximum encoded length for frame size, given we send less than 2^14 bytes

            let len = frame
                .data
                .len()
                .min(2usize.pow(14) - 1)
                .min(max_crypto_data_size);

            let data = frame.data.split_to(len);
            let offset = frame.offset;
            let truncated = frame::Crypto { offset, data };
            builder.write_frame(truncated, stats);

            if !frame.data.is_empty() {
                frame.offset += len as u64;
                space.pending.crypto.push_front(frame);
            }
        }

        // PATH_STATUS_AVAILABLE & PATH_STATUS_BACKUP
        while space_id == SpaceId::Data
            && !scheduling_info.is_abandoned
            && scheduling_info.may_send_data
            && frame::PathStatusAvailable::SIZE_BOUND <= builder.frame_space_remaining()
        {
            let Some(path_id) = space.pending.path_status.pop_first() else {
                break;
            };
            let Some(path) = self.paths.get(&path_id).map(|path_state| &path_state.data) else {
                trace!(%path_id, "discarding queued path status for unknown path");
                continue;
            };

            let seq = path.status.seq();
            match path.local_status() {
                PathStatus::Available => {
                    let frame = frame::PathStatusAvailable {
                        path_id,
                        status_seq_no: seq,
                    };
                    builder.write_frame(frame, stats);
                }
                PathStatus::Backup => {
                    let frame = frame::PathStatusBackup {
                        path_id,
                        status_seq_no: seq,
                    };
                    builder.write_frame(frame, stats);
                }
            }
        }

        // MAX_PATH_ID
        if space_id == SpaceId::Data
            && !scheduling_info.is_abandoned
            && scheduling_info.may_send_data
            && space.pending.max_path_id
            && frame::MaxPathId::SIZE_BOUND <= builder.frame_space_remaining()
        {
            let frame = frame::MaxPathId(self.local_max_path_id);
            builder.write_frame(frame, stats);
            space.pending.max_path_id = false;
        }

        // PATHS_BLOCKED
        if space_id == SpaceId::Data
            && !scheduling_info.is_abandoned
            && scheduling_info.may_send_data
            && frame::PathsBlocked::SIZE_BOUND <= builder.frame_space_remaining()
            && let Some(remote_max_path_id) = space.pending.paths_blocked.take()
        {
            let frame = frame::PathsBlocked(remote_max_path_id);
            builder.write_frame(frame, stats);
        }

        // PATH_CIDS_BLOCKED
        while space_id == SpaceId::Data
            && !scheduling_info.is_abandoned
            && scheduling_info.may_send_data
            && frame::PathCidsBlocked::SIZE_BOUND <= builder.frame_space_remaining()
        {
            let Some((path_id, next_seq)) = space.pending.path_cids_blocked.pop_first() else {
                break;
            };
            let frame = frame::PathCidsBlocked { path_id, next_seq };
            builder.write_frame(frame, stats);
        }

        // RESET_STREAM, STOP_SENDING, MAX_DATA, MAX_STREAM_DATA, MAX_STREAMS
        if space_id == SpaceId::Data
            && !scheduling_info.is_abandoned
            && scheduling_info.may_send_data
        {
            self.streams
                .write_control_frames(builder, &mut space.pending, stats);
        }

        // NEW_CONNECTION_ID
        let cid_len = self
            .local_cid_state
            .values()
            .map(|cid_state| cid_state.cid_len())
            .max()
            .expect("some local CID state must exist");
        let new_cid_size_bound =
            frame::NewConnectionId::size_bound(is_multipath_negotiated, cid_len);
        while !scheduling_info.is_abandoned
            && scheduling_info.may_send_data
            && builder.frame_space_remaining() > new_cid_size_bound
        {
            let Some(issued) = space.pending.new_cids.pop() else {
                break;
            };
            // Path was discarded after this CID was queued, drop.
            let Some(cid_state) = self.local_cid_state.get(&issued.path_id) else {
                debug!(
                    path = %issued.path_id, seq = issued.sequence,
                    "dropping queued NEW_CONNECTION_ID for discarded path",
                );
                continue;
            };
            let retire_prior_to = cid_state.retire_prior_to();

            let cid_path_id = match is_multipath_negotiated {
                true => Some(issued.path_id),
                false => {
                    debug_assert_eq!(issued.path_id, PathId::ZERO);
                    None
                }
            };
            let frame = frame::NewConnectionId {
                path_id: cid_path_id,
                sequence: issued.sequence,
                retire_prior_to,
                id: issued.id,
                reset_token: issued.reset_token,
            };
            builder.write_frame(frame, stats);
        }

        // RETIRE_CONNECTION_ID
        let retire_cid_bound = frame::RetireConnectionId::size_bound(is_multipath_negotiated);
        while !scheduling_info.is_abandoned
            && scheduling_info.may_send_data
            && builder.frame_space_remaining() > retire_cid_bound
        {
            let (path_id, sequence) = match space.pending.retire_cids.pop() {
                Some((PathId::ZERO, seq)) if !is_multipath_negotiated => (None, seq),
                Some((path_id, seq)) => (Some(path_id), seq),
                None => break,
            };
            let frame = frame::RetireConnectionId { path_id, sequence };
            builder.write_frame(frame, stats);
        }

        // DATAGRAM
        let mut sent_datagrams = false;
        while !scheduling_info.is_abandoned
            && scheduling_info.may_send_data
            && builder.frame_space_remaining() > Datagram::SIZE_BOUND
            && space_id == SpaceId::Data
        {
            match self.datagrams.write(builder, stats) {
                true => {
                    sent_datagrams = true;
                }
                false => break,
            }
        }
        if self.datagrams.send_blocked && sent_datagrams {
            self.events.push_back(Event::DatagramsUnblocked);
            self.datagrams.send_blocked = false;
        }

        let path = &mut self.paths.get_mut(&path_id).expect("known path").data;

        // NEW_TOKEN
        if !scheduling_info.is_abandoned && scheduling_info.may_send_data {
            while let Some(network_path) = space.pending.new_tokens.pop() {
                debug_assert_eq!(space_id, SpaceId::Data);
                let ConnectionSide::Server { server_config } = &self.side else {
                    panic!("NEW_TOKEN frames should not be enqueued by clients");
                };

                if !network_path.is_probably_same_path(&path.network_path) {
                    // NEW_TOKEN frames contain tokens bound to a client's IP address, and are only
                    // useful if used from the same IP address.  Thus, we abandon enqueued NEW_TOKEN
                    // frames upon an path change. Instead, when the new path becomes validated,
                    // NEW_TOKEN frames may be enqueued for the new path instead.
                    continue;
                }

                let token = Token::new(
                    TokenPayload::Validation {
                        ip: network_path.remote.ip(),
                        issued: server_config.time_source.now(),
                    },
                    &mut self.rng,
                );
                let new_token = NewToken {
                    token: token.encode(&*server_config.token_key).into(),
                };

                if builder.frame_space_remaining() < new_token.size() {
                    space.pending.new_tokens.push(network_path);
                    break;
                }

                builder.write_frame(new_token, stats);
                builder.retransmits_mut().new_tokens.push(network_path);
            }
        }

        // STREAM
        if !scheduling_info.is_abandoned
            && scheduling_info.may_send_data
            && space_id == SpaceId::Data
        {
            self.streams
                .write_stream_frames(builder, self.config.send_fairness, stats);
        }
    }

    /// Write pending ACKs into a buffer
    fn populate_acks<'a, 'b>(
        now: Instant,
        receiving_ecn: bool,
        path_id: PathId,
        space_id: SpaceId,
        space: &mut PacketSpace,
        is_multipath_negotiated: bool,
        builder: &mut PacketBuilder<'a, 'b>,
        stats: &mut FrameStats,
        space_has_keys: bool,
    ) {
        // 0-RTT packets must never carry acks (which would have to be of handshake packets)
        debug_assert!(space_has_keys, "tried to send ACK in 0-RTT");

        debug_assert!(
            is_multipath_negotiated || path_id == PathId::ZERO,
            "Only PathId::ZERO allowed without multipath (have {path_id:?})"
        );
        if is_multipath_negotiated {
            debug_assert!(
                space_id == SpaceId::Data || path_id == PathId::ZERO,
                "path acks must be sent in 1RTT space (have {space_id:?})"
            );
        }

        let pns = space.for_path(path_id);
        let ranges = pns.pending_acks.ranges();
        debug_assert!(!ranges.is_empty(), "can not send empty ACK range");
        let ecn = if receiving_ecn {
            Some(&pns.ecn_counters)
        } else {
            None
        };

        let delay_micros = pns.pending_acks.ack_delay(now).as_micros() as u64;
        // TODO: This should come from `TransportConfig` if that gets configurable.
        let ack_delay_exp = TransportParameters::default().ack_delay_exponent;
        let delay = delay_micros >> ack_delay_exp.into_inner();

        if is_multipath_negotiated && space_id == SpaceId::Data {
            if !ranges.is_empty() {
                let frame = frame::PathAck::encoder(path_id, delay, ranges, ecn);
                builder.write_frame(frame, stats);
            }
        } else {
            builder.write_frame(frame::Ack::encoder(delay, ranges, ecn), stats);
        }
    }

    fn close_common(&mut self) {
        trace!("connection closed");
        self.timers.reset();
    }

    fn set_close_timer(&mut self, now: Instant) {
        // QUIC-MULTIPATH § 2.6 Connection Closure: draining for 3*PTO using the max PTO of
        // all paths.
        let pto_max = self.max_pto_for_space(self.highest_space);
        self.timers.set(
            Timer::Conn(ConnTimer::Close),
            now + 3 * pto_max,
            self.qlog.with_time(now),
        );
    }

    /// Handle transport parameters received from the peer
    ///
    /// *remote_cid* and *local_cid* are the source and destination CIDs respectively of the
    /// *packet into which the transport parameters arrived.
    fn handle_peer_params(
        &mut self,
        params: TransportParameters,
        local_cid: ConnectionId,
        remote_cid: ConnectionId,
        now: Instant,
    ) -> Result<(), TransportError> {
        if Some(self.original_remote_cid) != params.initial_src_cid
            || (self.side.is_client()
                && (Some(self.initial_dst_cid) != params.original_dst_cid
                    || self.retry_src_cid != params.retry_src_cid))
        {
            return Err(TransportError::TRANSPORT_PARAMETER_ERROR(
                "CID authentication failure",
            ));
        }
        if params.initial_max_path_id.is_some() && (local_cid.is_empty() || remote_cid.is_empty()) {
            return Err(TransportError::PROTOCOL_VIOLATION(
                "multipath must not use zero-length CIDs",
            ));
        }

        self.set_peer_params(params);
        self.qlog.emit_peer_transport_params_received(self, now);

        Ok(())
    }

    fn set_peer_params(&mut self, params: TransportParameters) {
        self.streams.set_params(&params);
        self.idle_timeout =
            negotiate_max_idle_timeout(self.config.max_idle_timeout, Some(params.max_idle_timeout));
        trace!("negotiated max idle timeout {:?}", self.idle_timeout);

        if let Some(ref info) = params.preferred_address {
            // During the handshake PathId::ZERO exists.
            self.remote_cids.get_mut(&PathId::ZERO).expect("not yet abandoned").insert(frame::NewConnectionId {
                path_id: None,
                sequence: 1,
                id: info.connection_id,
                reset_token: info.stateless_reset_token,
                retire_prior_to: 0,
            })
            .expect(
                "preferred address CID is the first received, and hence is guaranteed to be legal",
            );
            let remote = self.path_data(PathId::ZERO).network_path.remote;
            self.set_reset_token(PathId::ZERO, remote, info.stateless_reset_token);
        }
        self.ack_frequency.peer_max_ack_delay = get_max_ack_delay(&params);

        let mut multipath_enabled = false;
        if let (Some(local_max_path_id), Some(remote_max_path_id)) = (
            self.config.get_initial_max_path_id(),
            params.initial_max_path_id,
        ) {
            // multipath is enabled, register the local and remote maximums
            self.local_max_path_id = local_max_path_id;
            self.remote_max_path_id = remote_max_path_id;
            let initial_max_path_id = local_max_path_id.min(remote_max_path_id);
            debug!(%initial_max_path_id, "multipath negotiated");
            multipath_enabled = true;
        }

        if let Some((max_locally_allowed_remote_addresses, max_remotely_allowed_remote_addresses)) =
            self.config
                .max_remote_nat_traversal_addresses
                .zip(params.max_remote_nat_traversal_addresses)
        {
            if multipath_enabled {
                let max_local_addresses = max_remotely_allowed_remote_addresses.get();
                let max_remote_addresses = max_locally_allowed_remote_addresses.get();
                self.n0_nat_traversal = n0_nat_traversal::State::new(
                    max_remote_addresses,
                    max_local_addresses,
                    self.side(),
                );
                debug!(
                    %max_remote_addresses, %max_local_addresses,
                    "n0's nat traversal negotiated"
                );
            } else {
                debug!("n0 nat traversal enabled for both endpoints, but multipath is missing")
            }
        }

        self.peer_params = params;
        let peer_max_udp_payload_size =
            u16::try_from(self.peer_params.max_udp_payload_size.into_inner()).unwrap_or(u16::MAX);
        let address_discovery_negotiated = self
            .config
            .address_discovery_role
            .should_report(&self.peer_params.address_discovery_role);

        let path = self.path_data_mut(PathId::ZERO);
        path.pending.observed_address = address_discovery_negotiated;
        path.mtud
            .on_peer_max_udp_payload_size_received(peer_max_udp_payload_size);
    }

    /// Decrypts a packet, returning the packet number on success
    fn decrypt_packet(
        &mut self,
        now: Instant,
        path_id: PathId,
        packet: &mut Packet,
    ) -> Result<Option<u64>, Option<TransportError>> {
        let result = self
            .crypto_state
            .decrypt_packet_body(packet, path_id, &self.spaces)?;

        let Some(result) = result else {
            return Ok(None);
        };

        if result.outgoing_key_update_acked
            && let Some(prev) = self.crypto_state.prev_crypto.as_mut()
        {
            prev.end_packet = Some((result.packet_number, now));
            self.set_key_discard_timer(now, packet.header.space());
        }

        if result.incoming_key_update {
            trace!("key update authenticated");
            self.crypto_state
                .update_keys(Some((result.packet_number, now)), true);
            self.set_key_discard_timer(now, packet.header.space());
        }

        Ok(Some(result.packet_number))
    }

    fn peer_supports_ack_frequency(&self) -> bool {
        self.peer_params.min_ack_delay.is_some()
    }

    /// Send an IMMEDIATE_ACK frame to the remote endpoint
    ///
    /// According to the spec, this will result in an error if the remote endpoint does not support
    /// the Acknowledgement Frequency extension
    pub(crate) fn immediate_ack(&mut self, path_id: PathId) {
        debug_assert_eq!(
            self.highest_space,
            SpaceKind::Data,
            "immediate ack must be written in the data space"
        );
        self.spaces[SpaceId::Data]
            .for_path(path_id)
            .pending_immediate_ack = true;
    }

    /// Decodes a packet, returning its decrypted payload, so it can be inspected in tests
    #[cfg(test)]
    pub(crate) fn decode_packet(&self, event: &ConnectionEvent) -> Option<Vec<u8>> {
        let ConnectionEventInner::Datagram(DatagramConnectionEvent {
            path_id,
            first_decode,
            remaining,
            ..
        }) = &event.0
        else {
            return None;
        };

        if remaining.is_some() {
            panic!("Packets should never be coalesced in tests");
        }

        let decrypted_header = self
            .crypto_state
            .unprotect_header(first_decode.clone(), self.peer_params.stateless_reset_token)?;

        let mut packet = decrypted_header.packet?;
        self.crypto_state
            .decrypt_packet_body(&mut packet, *path_id, &self.spaces)
            .ok()?;

        Some(packet.payload.to_vec())
    }

    /// The number of bytes of packets containing retransmittable frames that have not been
    /// acknowledged or declared lost.
    #[cfg(test)]
    pub(crate) fn bytes_in_flight(&self) -> u64 {
        // TODO(@divma): consider including for multipath?
        self.path_data(PathId::ZERO).in_flight.bytes
    }

    /// Number of bytes worth of non-ack-only packets that may be sent
    #[cfg(test)]
    pub(crate) fn congestion_window(&self) -> u64 {
        let path = self.path_data(PathId::ZERO);
        path.congestion
            .window()
            .saturating_sub(path.in_flight.bytes)
    }

    /// Whether no timers but keepalive, idle, rtt, pushnewcid, and key discard are running
    #[cfg(test)]
    pub(crate) fn is_idle(&self) -> bool {
        let current_timers = self.timers.values();
        current_timers
            .into_iter()
            .filter(|(timer, _)| {
                !matches!(
                    timer,
                    Timer::Conn(ConnTimer::KeepAlive)
                        | Timer::PerPath(_, PathTimer::PathKeepAlive)
                        | Timer::Conn(ConnTimer::PushNewCid)
                        | Timer::Conn(ConnTimer::KeyDiscard)
                )
            })
            .min_by_key(|(_, time)| *time)
            .is_none_or(|(timer, _)| {
                matches!(
                    timer,
                    Timer::Conn(ConnTimer::Idle) | Timer::PerPath(_, PathTimer::PathIdle)
                )
            })
    }

    /// Whether explicit congestion notification is in use on outgoing packets.
    #[cfg(test)]
    pub(crate) fn using_ecn(&self) -> bool {
        self.path_data(PathId::ZERO).sending_ecn
    }

    /// The number of received bytes in the current path
    #[cfg(test)]
    pub(crate) fn total_recvd(&self) -> u64 {
        self.path_data(PathId::ZERO).total_recvd
    }

    #[cfg(test)]
    pub(crate) fn active_local_cid_seq(&self) -> (u64, u64) {
        self.local_cid_state
            .get(&PathId::ZERO)
            .unwrap()
            .active_seq()
    }

    #[cfg(test)]
    #[track_caller]
    pub(crate) fn active_local_path_cid_seq(&self, path_id: u32) -> (u64, u64) {
        self.local_cid_state
            .get(&PathId(path_id))
            .unwrap()
            .active_seq()
    }

    /// Instruct the peer to replace previously issued CIDs by sending a NEW_CONNECTION_ID frame
    /// with updated `retire_prior_to` field set to `v`
    #[cfg(test)]
    pub(crate) fn rotate_local_cid(&mut self, v: u64, now: Instant) {
        let n = self
            .local_cid_state
            .get_mut(&PathId::ZERO)
            .unwrap()
            .assign_retire_seq(v);
        debug_assert!(!self.state.is_drained()); // requirement for endpoint_events
        self.endpoint_events
            .push_back(EndpointEventInner::NeedIdentifiers(PathId::ZERO, now, n));
    }

    /// Check the current active remote CID sequence for `PathId::ZERO`
    #[cfg(test)]
    pub(crate) fn active_remote_cid_seq(&self) -> u64 {
        self.remote_cids.get(&PathId::ZERO).unwrap().active_seq()
    }

    /// Returns the detected maximum udp payload size for the current path
    #[cfg(test)]
    pub(crate) fn path_mtu(&self, path_id: PathId) -> u16 {
        self.path_data(path_id).current_mtu()
    }

    /// Triggers path validation on all paths
    #[cfg(test)]
    pub(crate) fn trigger_path_validation(&mut self) {
        for path in self.paths.values_mut() {
            path.data.pending_challenge = true;
        }
    }

    /// Simulates a protocol violation error for test purposes.
    #[cfg(test)]
    pub fn simulate_protocol_violation(&mut self, now: Instant) {
        if !self.state.is_closed() {
            self.state
                .move_to_closed(TransportError::PROTOCOL_VIOLATION("simulated violation"));
            self.close_common();
            if !self.state.is_drained() {
                self.set_close_timer(now);
            }
            self.connection_close_pending = true;
        }
    }

    /// Whether we have **on-path** 1-RTT data to send.
    ///
    /// This checks for frames that can only be sent in the data space (1-RTT):
    /// - Pending PATH_CHALLENGE frames on the active and previous path if just migrated.
    /// - Pending PATH_RESPONSE frames.
    /// - Pending data to send in STREAM frames.
    /// - Pending DATAGRAM frames to send.
    ///
    /// See also [`PacketSpace::can_send`] which keeps track of all other frame types that
    /// may need to be sent.
    fn can_send_1rtt(&self, path_id: PathId, max_size: usize) -> SendableFrames {
        let network_path = self.path_data(path_id).network_path;
        let space_specific = self
            .paths
            .get(&path_id)
            .is_some_and(|path| path.data.pending_challenge || !path.data.pending.is_empty())
            || self.spaces[SpaceKind::Data]
                .number_spaces
                .get(&path_id)
                .is_some_and(|pns| pns.pending_path_responses.has_pending_on_path(network_path));

        // Stream control frames are checked in PacketSpace::can_send, only check data here.
        let other = self.streams.can_send_stream_data()
            || self
                .datagrams
                .outgoing
                .front()
                .is_some_and(|x| x.size(true) <= max_size);

        // All `false` fields are set in PacketSpace::can_send.
        SendableFrames {
            acks: false,
            close: false,
            space_specific,
            other,
        }
    }

    /// Terminate the connection instantly, without sending a close packet
    fn kill(&mut self, reason: ConnectionError) {
        self.close_common();
        self.state
            .move_to_drained(Some(reason), &mut self.endpoint_events);
    }

    /// Storage size required for the largest packet that can be transmitted on all currently
    /// available paths
    ///
    /// Buffers passed to [`Connection::poll_transmit`] should be at least this large.
    ///
    /// When multipath is enabled, this value is the minimum MTU across all available paths.
    pub fn current_mtu(&self) -> u16 {
        self.paths
            .iter()
            .filter(|&(path_id, _path_state)| !self.abandoned_paths.contains(path_id))
            .map(|(_path_id, path_state)| path_state.data.current_mtu())
            .min()
            .unwrap_or(INITIAL_MTU)
    }

    /// Size of non-frame data for a 1-RTT packet
    ///
    /// Quantifies space consumed by the QUIC header and AEAD tag. All other bytes in a packet are
    /// frames. Changes if the length of the remote connection ID changes, which is expected to be
    /// rare. If `pn` is specified, may additionally change unpredictably due to variations in
    /// latency and packet loss.
    fn predict_1rtt_overhead(&mut self, pn: u64, path: PathId) -> usize {
        let pn_len = PacketNumber::new(
            pn,
            self.spaces[SpaceId::Data]
                .for_path(path)
                .largest_acked_packet_pn
                .unwrap_or(0),
        )
        .len();

        // 1 byte for flags
        1 + self
            .remote_cids
            .get(&path)
            .map(|cids| cids.active().len())
            .unwrap_or(20)      // Max CID len in QUIC v1
            + pn_len
            + self.tag_len_1rtt()
    }

    fn predict_1rtt_overhead_no_pn(&self) -> usize {
        let pn_len = 4;

        let cid_len = self
            .remote_cids
            .values()
            .map(|cids| cids.active().len())
            .max()
            .unwrap_or(20); // Max CID len in QUIC v1

        // 1 byte for flags
        1 + cid_len + pn_len + self.tag_len_1rtt()
    }

    fn tag_len_1rtt(&self) -> usize {
        // encryption_keys for Data space returns 1-RTT keys if available, otherwise 0-RTT keys
        let packet_crypto = self
            .crypto_state
            .encryption_keys(SpaceKind::Data, self.side.side())
            .map(|(_header, packet, _level)| packet);
        // If neither Data nor 0-RTT keys are available, make a reasonable tag length guess. As of
        // this writing, all QUIC cipher suites use 16-byte tags. We could return `None` instead,
        // but that would needlessly prevent sending datagrams during 0-RTT.
        packet_crypto.map_or(16, |x| x.tag_len())
    }

    /// Mark the path as validated, and enqueue NEW_TOKEN frames to be sent as appropriate
    fn on_path_validated(&mut self, path_id: PathId) {
        self.path_data_mut(path_id).validated = true;
        let ConnectionSide::Server { server_config } = &self.side else {
            return;
        };
        let network_path = self.path_data(path_id).network_path;
        let new_tokens = &mut self.spaces[SpaceId::Data as usize].pending.new_tokens;
        new_tokens.clear();
        for _ in 0..server_config.validation_token.sent {
            new_tokens.push(network_path);
        }
    }

    /// Handle new path status information: PATH_STATUS_AVAILABLE, PATH_STATUS_BACKUP
    fn on_path_status(&mut self, path_id: PathId, status: PathStatus, status_seq_no: VarInt) {
        if let Some(path) = self.paths.get_mut(&path_id) {
            path.data.status.remote_update(status, status_seq_no);
        } else {
            debug!("PATH_STATUS_AVAILABLE received unknown path {:?}", path_id);
        }
        self.events.push_back(
            PathEvent::RemoteStatus {
                id: path_id,
                status,
            }
            .into(),
        );
    }

    /// Returns the maximum [`PathId`] to be used for sending in this connection.
    ///
    /// This is calculated as minimum between the local and remote's maximums when multipath is
    /// enabled, or `None` when disabled.
    ///
    /// For data that's received, we should use [`Self::local_max_path_id`] instead.
    /// The reasoning is that the remote might already have updated to its own newer
    /// [`Self::max_path_id`] after sending out a `MAX_PATH_ID` frame, but it got re-ordered.
    fn max_path_id(&self) -> Option<PathId> {
        if self.is_multipath_negotiated() {
            Some(self.remote_max_path_id.min(self.local_max_path_id))
        } else {
            None
        }
    }

    /// Returns whether this connection has a socket that supports IPv6.
    ///
    /// TODO(matheus23): This is related to noq endpoint state's `ipv6` bool. We should move that info
    /// here instead of trying to hack around not knowing it exactly.
    pub(crate) fn is_ipv6(&self) -> bool {
        self.paths
            .values()
            .any(|p| p.data.network_path.remote.is_ipv6())
    }

    /// Add addresses the local endpoint considers are reachable for nat traversal.
    pub fn add_nat_traversal_address(
        &mut self,
        address: SocketAddr,
    ) -> Result<(), n0_nat_traversal::Error> {
        if let Some(added) = self.n0_nat_traversal.add_local_address(address)? {
            self.spaces[SpaceId::Data].pending.add_address.insert(added);
        };
        Ok(())
    }

    /// Removes an address the endpoing no longer considers reachable for nat traversal
    ///
    /// Addresses not present in the set will be silently ignored.
    pub fn remove_nat_traversal_address(
        &mut self,
        address: SocketAddr,
    ) -> Result<(), n0_nat_traversal::Error> {
        if let Some(removed) = self.n0_nat_traversal.remove_local_address(address)? {
            self.spaces[SpaceId::Data]
                .pending
                .remove_address
                .insert(removed);
        }
        Ok(())
    }

    /// Get the current local nat traversal addresses
    pub fn get_local_nat_traversal_addresses(
        &self,
    ) -> Result<Vec<SocketAddr>, n0_nat_traversal::Error> {
        self.n0_nat_traversal.get_local_nat_traversal_addresses()
    }

    /// Get the currently advertised nat traversal addresses by the server
    pub fn get_remote_nat_traversal_addresses(
        &self,
    ) -> Result<Vec<SocketAddr>, n0_nat_traversal::Error> {
        Ok(self
            .n0_nat_traversal
            .client_side()?
            .get_remote_nat_traversal_addresses())
    }

    /// Initiates a new nat traversal round
    ///
    /// A nat traversal round involves advertising the client's local addresses in
    /// `REACH_OUT` frames, and initiating probing of the known remote addresses. When a new
    /// round is initiated, the previous one is cancelled.
    ///
    /// For all probes that succeed, if any, a new path will be opened on the successful
    /// 4-tuple.
    ///
    /// Returns the server addresses that are now being probed. If addresses fail due to
    /// spurious errors, these might succeed later and not be returned in this set.
    pub fn initiate_nat_traversal_round(
        &mut self,
        now: Instant,
    ) -> Result<Vec<SocketAddr>, n0_nat_traversal::Error> {
        if self.state.is_closed() {
            return Err(n0_nat_traversal::Error::Closed);
        }

        let ipv6 = self.is_ipv6();
        let client_state = self.n0_nat_traversal.client_side_mut()?;
        let (mut reach_out_frames, probed_addrs) =
            client_state.initiate_nat_traversal_round(ipv6)?;
        if let Some(delay) = self.n0_nat_traversal.retry_delay(self.config.initial_rtt) {
            self.timers.set(
                Timer::Conn(ConnTimer::NatTraversalProbeRetry),
                now + delay,
                self.qlog.with_time(now),
            );
        }

        self.spaces[SpaceId::Data]
            .pending
            .reach_out
            .append(&mut reach_out_frames);

        Ok(probed_addrs)
    }

    /// Whether the handshake is considered **confirmed**.
    ///
    /// <https://www.rfc-editor.org/rfc/rfc9001#section-4.1.2> defines a handshake to be
    /// confirmed when you know the peer successfully received and successfully processed
    /// your TLS Finished message.
    ///
    /// Implementation-wise this is the point at which the handshake crypto keys are
    /// discarded. So we can use this to know if the handshake is confirmed.
    fn is_handshake_confirmed(&self) -> bool {
        !self.is_handshaking() && !self.crypto_state.has_keys(EncryptionLevel::Handshake)
    }
}

impl fmt::Debug for Connection {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        f.debug_struct("Connection")
            .field("handshake_cid", &self.handshake_cid)
            .finish()
    }
}

/// The set of abandoned paths.
///
/// Implementation based on ArrayRangeSet to share more code. The memory space is
/// proportional to the number of concurrently open paths allowed. So does not grow
/// unbounded.
#[derive(Debug, Default)]
struct AbandonedPaths(ArrayRangeSet<ABANDONED_PATH_INLINE_RANGES, u32>);

/// Size of the stack-allocated array in the [`ArrayRangeSet`].
///
/// A range is 2 u32's, so this is 16 * (4 + 4) = 128 bytes. A good size for inline data,
/// with plenty of ranges for common multipath use.
const ABANDONED_PATH_INLINE_RANGES: usize = 16;

impl AbandonedPaths {
    /// The number of abandoned paths.
    fn len(&self) -> u32 {
        self.0.elts_count()
    }

    /// The largest abandoned path.
    fn max(&self) -> Option<PathId> {
        self.0.max().map(PathId::from)
    }

    /// Whether the path is already abandoned.
    fn contains(&self, val: &PathId) -> bool {
        self.0.contains(val.as_u32())
    }

    /// Adds another abandoned path.
    fn insert(&mut self, val: PathId) {
        self.0.insert_one(val.as_u32());
    }
}

/// Hints when the caller identifies a network change.
pub trait NetworkChangeHint: fmt::Debug + 'static {
    /// Inform the connection if a path may recover after a network change.
    ///
    /// After network changes, paths may not be recoverable. In this case, waiting for the path to
    /// become idle may take longer than what is desirable. If [`Self::is_path_recoverable`]
    /// returns `false`, a multipath-enabled, client-side connection will establish a new path to
    /// the same remote, closing the current one, instead of migrating the path.
    ///
    /// Paths that are deemed recoverable will simply be sent a PING for a liveness check.
    fn is_path_recoverable(&self, path_id: PathId, network_path: FourTuple) -> bool;
}

/// Return value for [`Connection::poll_transmit_path_space`].
#[derive(Debug)]
enum PollPathSpaceStatus {
    /// Nothing to send in the space, nothing was written into the [`TransmitBuf`].
    NothingToSend {
        /// If true there was data to send but congestion control did not allow so.
        congestion_blocked: bool,
    },
    /// One or more packets have been written into the [`TransmitBuf`].
    WrotePacket {
        /// The highest packet number.
        last_packet_number: u64,
        /// Whether to pad an already started datagram in the next packet.
        ///
        /// When packets in Initial, 0-RTT or Handshake packet do not fill the entire
        /// datagram they may decide to coalesce with the next packet from a higher
        /// encryption level on the same path. But the earlier packet may require specific
        /// size requirements for the datagram they are sent in.
        ///
        /// If a space did not complete the datagram, they use this to request the correct
        /// padding in the final packet of the datagram so that the final datagram will have
        /// the correct size.
        ///
        /// If a space did fill an entire datagram, it leaves this to the default of
        /// [`PadDatagram::No`].
        pad_datagram: PadDatagram,
    },
    /// Send the contents of the transmit immediately.
    ///
    /// Packets were written and the GSO batch must end now, regardless from whether higher
    /// spaces still have frames to write. This is used when the last datagram written would
    /// require too much padding to continue a GSO batch, which would waste space on the
    /// wire.
    Send {
        /// The highest packet number written into the transmit.
        last_packet_number: u64,
    },
}

/// Information used to decide what frames to schedule into which packets.
///
/// Primarily used by [`Connection::poll_transmit_on_path`] and the functions that help
/// building packets for it: [`Connection::poll_transmit_path_space`] and
/// [`Connection::populate_packet`].
#[derive(Debug, Copy, Clone)]
struct PathSchedulingInfo {
    /// Whether the path is abandoned.
    ///
    /// Note that a path that is abandoned but still has CIDs can still send a packet. After
    /// sending that packet the CIDs issued by the remote have to be considered retired as
    /// well.
    is_abandoned: bool,
    /// Whether the path may send [`SpaceKind::Data`] frames.
    ///
    /// Some paths should only send frames from [`SendableFrames::space_specific`]. All other
    /// frames are essentially frames that can be sent on any [`SpaceKind::Data`] space. For
    /// those we want to respect packet scheduling rules however.
    ///
    /// Roughly speaking data frames are only sent on spaces that have CIDs, are not
    /// abandoned and have no *better* spaces. However see to comments where this is
    /// populated for the exact packet scheduling implementation.
    ///
    /// This essentially marks this paths as the best validated space ID. Except during
    /// the handshake in which case it does not need to be validated. Several paths could be
    /// equally good and all have this set to `true`, in that case packet scheduling can
    /// choose which path to use. Currently it chooses the lowest path that is not
    /// congestion blocked.
    ///
    /// Note that once in the closed or draining states this will never be true.
    may_send_data: bool,
    /// Whether the path may send a CONNECTION_CLOSE frame.
    ///
    /// This essentially marks this path as the best validated space ID with a fallback
    /// to unvalidated spaces if there are no validated spaces. Like for
    /// [`Self::may_send_data`] other paths could be equally good.
    may_send_close: bool,
    may_self_abandon: bool,
}

#[derive(Debug, Copy, Clone, PartialEq, Eq)]
enum PathBlocked {
    No,
    AntiAmplification,
    Congestion,
    Pacing,
}

/// Fields of `Connection` specific to it being client-side or server-side
enum ConnectionSide {
    Client {
        /// Sent in every outgoing Initial packet. Always empty after Initial keys are discarded
        token: Bytes,
        token_store: Arc<dyn TokenStore>,
        server_name: String,
    },
    Server {
        server_config: Arc<ServerConfig>,
    },
}

impl ConnectionSide {
    fn is_client(&self) -> bool {
        self.side().is_client()
    }

    fn is_server(&self) -> bool {
        self.side().is_server()
    }

    fn side(&self) -> Side {
        match *self {
            Self::Client { .. } => Side::Client,
            Self::Server { .. } => Side::Server,
        }
    }
}

impl From<SideArgs> for ConnectionSide {
    fn from(side: SideArgs) -> Self {
        match side {
            SideArgs::Client {
                token_store,
                server_name,
            } => Self::Client {
                token: token_store.take(&server_name).unwrap_or_default(),
                token_store,
                server_name,
            },
            SideArgs::Server {
                server_config,
                pref_addr_cid: _,
                path_validated: _,
            } => Self::Server { server_config },
        }
    }
}

/// Parameters to `Connection::new` specific to it being client-side or server-side
pub(crate) enum SideArgs {
    Client {
        token_store: Arc<dyn TokenStore>,
        server_name: String,
    },
    Server {
        server_config: Arc<ServerConfig>,
        pref_addr_cid: Option<ConnectionId>,
        path_validated: bool,
    },
}

impl SideArgs {
    pub(crate) fn pref_addr_cid(&self) -> Option<ConnectionId> {
        match *self {
            Self::Client { .. } => None,
            Self::Server { pref_addr_cid, .. } => pref_addr_cid,
        }
    }

    pub(crate) fn path_validated(&self) -> bool {
        match *self {
            Self::Client { .. } => true,
            Self::Server { path_validated, .. } => path_validated,
        }
    }

    pub(crate) fn side(&self) -> Side {
        match *self {
            Self::Client { .. } => Side::Client,
            Self::Server { .. } => Side::Server,
        }
    }
}

/// Reasons why a connection might be lost
#[derive(Debug, Error, Clone, PartialEq, Eq)]
pub enum ConnectionError {
    /// The peer doesn't implement any supported version
    #[error("peer doesn't implement any supported version")]
    VersionMismatch,
    /// The peer violated the QUIC specification as understood by this implementation
    #[error(transparent)]
    TransportError(#[from] TransportError),
    /// The peer's QUIC stack aborted the connection automatically
    #[error("aborted by peer: {0}")]
    ConnectionClosed(frame::ConnectionClose),
    /// The peer closed the connection
    #[error("closed by peer: {0}")]
    ApplicationClosed(frame::ApplicationClose),
    /// The peer is unable to continue processing this connection, usually due to having restarted
    #[error("reset by peer")]
    Reset,
    /// Communication with the peer has lapsed for longer than the negotiated idle timeout
    ///
    /// If neither side is sending keep-alives, a connection will time out after a long enough idle
    /// period even if the peer is still reachable. See also [`TransportConfig::max_idle_timeout()`]
    /// and [`TransportConfig::keep_alive_interval()`].
    #[error("timed out")]
    TimedOut,
    /// The local application closed the connection
    #[error("closed")]
    LocallyClosed,
    /// The connection could not be created because not enough of the CID space is available
    ///
    /// Try using longer connection IDs.
    #[error("CIDs exhausted")]
    CidsExhausted,
}

impl From<Close> for ConnectionError {
    fn from(x: Close) -> Self {
        match x {
            Close::Connection(reason) => Self::ConnectionClosed(reason),
            Close::Application(reason) => Self::ApplicationClosed(reason),
        }
    }
}

// For compatibility with API consumers
impl From<ConnectionError> for io::Error {
    fn from(x: ConnectionError) -> Self {
        use ConnectionError::*;
        let kind = match x {
            TimedOut => io::ErrorKind::TimedOut,
            Reset => io::ErrorKind::ConnectionReset,
            ApplicationClosed(_) | ConnectionClosed(_) => io::ErrorKind::ConnectionAborted,
            TransportError(_) | VersionMismatch | LocallyClosed | CidsExhausted => {
                io::ErrorKind::Other
            }
        };
        Self::new(kind, x)
    }
}

/// Errors that might trigger a path being closed
// TODO(@divma): maybe needs to be reworked based on what we want to do with the public API
#[derive(Debug, Error, PartialEq, Eq, Clone, Copy)]
pub enum PathError {
    /// The extension was not negotiated with the peer
    #[error("multipath extension not negotiated")]
    MultipathNotNegotiated,
    /// Paths can only be opened client-side
    #[error("the server side may not open a path")]
    ServerSideNotAllowed,
    /// Current limits do not allow us to open more paths
    #[error("maximum number of concurrent paths reached")]
    MaxPathIdReached,
    /// No remote CIDs available to open a new path
    #[error("remoted CIDs exhausted")]
    RemoteCidsExhausted,
    /// Path could not be validated and will be abandoned
    #[error("path validation failed")]
    ValidationFailed,
    /// The remote address for the path is not supported by the endpoint
    #[error("invalid remote address")]
    InvalidRemoteAddress(SocketAddr),
}

/// Errors triggered when abandoning a path
#[derive(Debug, Error, Clone, Eq, PartialEq)]
pub enum ClosePathError {
    /// Multipath is not negotiated
    #[error("Multipath extension not negotiated")]
    MultipathNotNegotiated,
    /// The path is already closed or was never opened
    #[error("closed path")]
    ClosedPath,
    /// Cannot close the last remaining open path via the local API.
    ///
    /// Use [`Connection::close`] to end the connection instead.
    #[error("last open path")]
    LastOpenPath,
}

/// Error when the multipath extension was not negotiated, but attempted to be used.
#[derive(Debug, Error, Clone, Copy)]
#[error("Multipath extension not negotiated")]
pub struct MultipathNotNegotiated {
    _private: (),
}

/// Events of interest to the application
#[derive(Debug)]
pub enum Event {
    /// The connection's handshake data is ready
    HandshakeDataReady,
    /// The connection was successfully established
    Connected,
    /// The TLS handshake was confirmed
    HandshakeConfirmed,
    /// The connection was lost
    ///
    /// Emitted when the connection is closed due to an error, a timeout, or the peer closing it.
    /// This is **not** emitted when the local application closes the connection via
    /// [`Connection::close()`](crate::Connection::close). In that case, pending operations will
    /// fail with [`ConnectionError::LocallyClosed`].
    ConnectionLost {
        /// Reason that the connection was closed
        reason: ConnectionError,
    },
    /// Stream events
    Stream(StreamEvent),
    /// One or more application datagrams have been received
    DatagramReceived,
    /// One or more application datagrams have been sent after blocking
    DatagramsUnblocked,
    /// (Multi)Path events
    Path(PathEvent),
    /// n0's nat traversal events
    NatTraversal(n0_nat_traversal::Event),
}

impl From<PathEvent> for Event {
    fn from(source: PathEvent) -> Self {
        Self::Path(source)
    }
}

fn get_max_ack_delay(params: &TransportParameters) -> Duration {
    Duration::from_micros(params.max_ack_delay.0 * 1000)
}

/// Prevents overflow and improves behavior in extreme circumstances.
const MAX_BACKOFF_EXPONENT: u32 = 16;

/// The max interval between successive tail-loss probes.
///
/// This is the "normal" value we use.
const MAX_PTO_INTERVAL: Duration = Duration::from_secs(2);

/// The idle time, below which we use the shorter [`MAX_PTO_FAST_INTERVAL`].
const MIN_IDLE_FOR_FAST_PTO: Duration = Duration::from_secs(25);

/// The max interval between successive tail-loss probes with short idle times.
///
/// If the path or connection idle time is less than [`MIN_IDLE_FOR_FAST_PTO`] then we use
/// this value to ensure we have plenty of retransmits before we reach the idle time.
const MAX_PTO_FAST_INTERVAL: Duration = Duration::from_secs(1);

/// The RTT threshold above which we cap the PTO interval to 1.5 * smoothed_rtt
///
/// This is RTT time above which 1.5 * RTT > [`MAX_PTO_INTERVAL`], for these links we want
/// to extend the interval between tail-loss probes to not fill the entire pipe with them.
const SLOW_RTT_THRESHOLD: Duration =
    Duration::from_millis((MAX_PTO_INTERVAL.as_millis() as u64 * 2) / 3);

/// Minimal remaining size to allow packet coalescing, excluding cryptographic tag
///
/// This must be at least as large as the header for a well-formed empty packet to be coalesced,
/// plus some space for frames. We only care about handshake headers because short header packets
/// necessarily have smaller headers, and initial packets are only ever the first packet in a
/// datagram (because we coalesce in ascending packet space order and the only reason to split a
/// packet is when packet space changes).
const MIN_PACKET_SPACE: usize = MAX_HANDSHAKE_OR_0RTT_HEADER_SIZE + 32;

/// Largest amount of space that could be occupied by a Handshake or 0-RTT packet's header
///
/// Excludes packet-type-specific fields such as packet number or Initial token
// https://www.rfc-editor.org/rfc/rfc9000.html#name-0-rtt: flags + version + dcid len + dcid +
// scid len + scid + length + pn
const MAX_HANDSHAKE_OR_0RTT_HEADER_SIZE: usize =
    1 + 4 + 1 + MAX_CID_SIZE + 1 + MAX_CID_SIZE + VarInt::from_u32(u16::MAX as u32).size() + 4;

#[derive(Default)]
struct SentFrames {
    retransmits: ThinRetransmits,
    path_retransmits: PathRetransmits,
    /// The packet number of the largest acknowledged packet for each path
    largest_acked: FxHashMap<PathId, u64>,
    stream_frames: StreamMetaVec,
    /// Whether the packet contains non-retransmittable frames (like datagrams)
    non_retransmits: bool,
    /// If the datagram containing these frames should be padded to the min MTU
    requires_padding: bool,
}

impl SentFrames {
    /// Returns whether the packet contains only ACKs
    fn is_ack_only(&self, streams: &StreamsState) -> bool {
        !self.largest_acked.is_empty()
            && !self.non_retransmits
            && self.stream_frames.is_empty()
            && self.retransmits.is_empty(streams)
    }

    fn retransmits_mut(&mut self) -> &mut Retransmits {
        self.retransmits.get_or_create()
    }

    fn record_sent_frame(&mut self, frame: frame::EncodableFrame<'_>) {
        use frame::EncodableFrame::*;
        match frame {
            PathAck(path_ack_encoder) => {
                if let Some(max) = path_ack_encoder.ranges.max() {
                    self.largest_acked.insert(path_ack_encoder.path_id, max);
                }
            }
            Ack(ack_encoder) => {
                if let Some(max) = ack_encoder.ranges.max() {
                    self.largest_acked.insert(PathId::ZERO, max);
                }
            }
            Close(_) => { /* non retransmittable, but after this we don't really care */ }
            PathResponse(_) => self.non_retransmits = true,
            HandshakeDone(_) => self.retransmits_mut().handshake_done = true,
            ReachOut(frame) => self.retransmits_mut().reach_out.push(frame),
            ObservedAddr(_) => self.path_retransmits.observed_address = true,
            Ping(_) => self.non_retransmits = true,
            ImmediateAck(_) => self.non_retransmits = true,
            AckFrequency(_) => self.retransmits_mut().ack_frequency = true,
            PathChallenge(_) => self.non_retransmits = true,
            Crypto(crypto) => self.retransmits_mut().crypto.push_back(crypto),
            PathAbandon(path_abandon) => {
                self.retransmits_mut()
                    .path_abandon
                    .entry(path_abandon.path_id)
                    .or_insert(path_abandon.error_code);
            }
            PathStatusAvailable(frame::PathStatusAvailable { path_id, .. })
            | PathStatusBackup(frame::PathStatusBackup { path_id, .. }) => {
                self.retransmits_mut().path_status.insert(path_id);
            }
            MaxPathId(_) => self.retransmits_mut().max_path_id = true,
            PathsBlocked(frame::PathsBlocked(path_id)) => {
                let paths_blocked = &mut self.retransmits_mut().paths_blocked;
                *paths_blocked = cmp::max(*paths_blocked, Some(path_id));
            }
            PathCidsBlocked(path_cids_blocked) => {
                self.retransmits_mut()
                    .path_cids_blocked
                    .entry(path_cids_blocked.path_id)
                    .and_modify(|next_seq| {
                        *next_seq = cmp::max(*next_seq, path_cids_blocked.next_seq);
                    })
                    .or_insert(path_cids_blocked.next_seq);
            }
            ResetStream(reset) => self
                .retransmits_mut()
                .reset_stream
                .push((reset.id, reset.error_code)),
            StopSending(stop_sending) => self.retransmits_mut().stop_sending.push(stop_sending),
            NewConnectionId(new_cid) => self.retransmits_mut().new_cids.push(new_cid.issued()),
            RetireConnectionId(retire_cid) => self
                .retransmits_mut()
                .retire_cids
                .push((retire_cid.path_id.unwrap_or_default(), retire_cid.sequence)),
            Datagram(_) => self.non_retransmits = true,
            NewToken(_) => {}
            AddAddress(add_address) => {
                self.retransmits_mut().add_address.insert(add_address);
            }
            RemoveAddress(remove_address) => {
                self.retransmits_mut().remove_address.insert(remove_address);
            }
            StreamMeta(stream_meta_encoder) => self.stream_frames.push(stream_meta_encoder.meta),
            MaxData(_) => self.retransmits_mut().max_data = true,
            MaxStreamData(max) => {
                self.retransmits_mut().max_stream_data.insert(max.id);
            }
            MaxStreams(max_streams) => {
                self.retransmits_mut().max_stream_id[max_streams.dir as usize] = true
            }
            StreamsBlocked(streams_blocked) => {
                self.retransmits_mut().streams_blocked[streams_blocked.dir as usize] = true
            }
        }
    }
}

/// Compute the negotiated idle timeout based on local and remote max_idle_timeout transport parameters.
///
/// According to the definition of max_idle_timeout, a value of `0` means the timeout is disabled; see <https://www.rfc-editor.org/rfc/rfc9000#section-18.2-4.4.1.>
///
/// According to the negotiation procedure, either the minimum of the timeouts or one specified is used as the negotiated value; see <https://www.rfc-editor.org/rfc/rfc9000#section-10.1-2.>
///
/// Returns the negotiated idle timeout as a `Duration`, or `None` when both endpoints have opted out of idle timeout.
fn negotiate_max_idle_timeout(x: Option<VarInt>, y: Option<VarInt>) -> Option<Duration> {
    match (x, y) {
        (Some(VarInt(0)) | None, Some(VarInt(0)) | None) => None,
        (Some(VarInt(0)) | None, Some(y)) => Some(Duration::from_millis(y.0)),
        (Some(x), Some(VarInt(0)) | None) => Some(Duration::from_millis(x.0)),
        (Some(x), Some(y)) => Some(Duration::from_millis(cmp::min(x, y).0)),
    }
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn negotiate_max_idle_timeout_commutative() {
        let test_params = [
            (None, None, None),
            (None, Some(VarInt(0)), None),
            (None, Some(VarInt(2)), Some(Duration::from_millis(2))),
            (Some(VarInt(0)), Some(VarInt(0)), None),
            (
                Some(VarInt(2)),
                Some(VarInt(0)),
                Some(Duration::from_millis(2)),
            ),
            (
                Some(VarInt(1)),
                Some(VarInt(4)),
                Some(Duration::from_millis(1)),
            ),
        ];

        for (left, right, result) in test_params {
            assert_eq!(negotiate_max_idle_timeout(left, right), result);
            assert_eq!(negotiate_max_idle_timeout(right, left), result);
        }
    }

    #[test]
    fn abandoned_paths() {
        let mut t = AbandonedPaths::default();

        t.insert(PathId(0));
        t.insert(PathId(1));
        assert_eq!(t.len(), 2);
        assert_eq!(t.0.range_count(), 1); // 2 elements compacted into one range
        assert!(t.contains(&PathId(0)));
        assert!(t.contains(&PathId(1)));
        assert!(!t.contains(&PathId(2)));
        assert!(!t.contains(&PathId(3)));
        assert_eq!(t.max(), Some(PathId(1)));

        t.insert(PathId(3));
        assert_eq!(t.len(), 3);
        assert_eq!(t.0.range_count(), 2); // 3 elements compacted into 2 ranges
        assert!(t.contains(&PathId(0)));
        assert!(t.contains(&PathId(1)));
        assert!(!t.contains(&PathId(2)));
        assert!(t.contains(&PathId(3)));
        assert_eq!(t.max(), Some(PathId(3)));

        t.insert(PathId(2));
        assert_eq!(t.len(), 4);
        assert_eq!(t.0.range_count(), 1); // 4 elements compacted into 1 range
        assert!(t.contains(&PathId(0)));
        assert!(t.contains(&PathId(1)));
        assert!(t.contains(&PathId(2)));
        assert!(t.contains(&PathId(3)));
        assert_eq!(t.max(), Some(PathId(3)));
    }
}