RFC 9002

QUIC Loss Detection and Congestion Control

📅 published May 2021, from ietf

RFC 9002: QUIC Loss Detection and Congestion Control 🦋

Okay bestie, so this is RFC 9002, a Standards Track document from the IETF (Internet Engineering Task Force) 💅. It came out in May 2021, and its ISSN is 2070-1721. Two editors wrote it: J. Iyengar, Ed. from Fastly and I. Swett, Ed. from Google 🫶. The header doesn't list anything it obsoletes or updates, so it's standing on its own, no cap 👑.

Abstract ✨

This doc describes the loss detection and congestion control mechanisms for QUIC. Basically, it's the rulebook for how QUIC notices packets that went missing and how it manages congestion 🌊

Status of This Memo 📜

This is an Internet Standards Track document.

This document is a product of the Internet Engineering Task Force (IETF). It represents the consensus of the IETF community. It has received public review and has been approved for publication by the Internet Engineering Steering Group (IESG). Further information on Internet Standards is available in Section 2 of RFC 7841.

Information about the current status of this document, any errata, and how to provide feedback on it may be obtained at https://www.rfc-editor.org/info/rfc9002.

Copyright Notice ⚖️

This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Simplified BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Simplified BSD License.

Table of Contents 🗂️

1.  Introduction
2.  Conventions and Definitions
3.  Design of the QUIC Transmission Machinery
4.  Relevant Differences between QUIC and TCP
  4.1.  Separate Packet Number Spaces
  4.2.  Monotonically Increasing Packet Numbers
  4.3.  Clearer Loss Epoch
  4.4.  No Reneging
  4.5.  More ACK Ranges
  4.6.  Explicit Correction for Delayed Acknowledgments
  4.7.  Probe Timeout Replaces RTO and TLP
  4.8.  The Minimum Congestion Window Is Two Packets
  4.9.  Handshake Packets Are Not Special
5.  Estimating the Round-Trip Time
  5.1.  Generating RTT Samples
  5.2.  Estimating min_rtt
  5.3.  Estimating smoothed_rtt and rttvar
6.  Loss Detection
  6.1.  Acknowledgment-Based Detection
    6.1.1.  Packet Threshold
    6.1.2.  Time Threshold
  6.2.  Probe Timeout
    6.2.1.  Computing PTO
    6.2.2.  Handshakes and New Paths
    6.2.3.  Speeding up Handshake Completion
    6.2.4.  Sending Probe Packets
  6.3.  Handling Retry Packets
  6.4.  Discarding Keys and Packet State
7.  Congestion Control
  7.1.  Explicit Congestion Notification
  7.2.  Initial and Minimum Congestion Window
  7.3.  Congestion Control States
    7.3.1.  Slow Start
    7.3.2.  Recovery
    7.3.3.  Congestion Avoidance
  7.4.  Ignoring Loss of Undecryptable Packets
  7.5.  Probe Timeout
  7.6.  Persistent Congestion
    7.6.1.  Duration
    7.6.2.  Establishing Persistent Congestion
    7.6.3.  Example
  7.7.  Pacing
  7.8.  Underutilizing the Congestion Window
8.  Security Considerations
  8.1.  Loss and Congestion Signals
  8.2.  Traffic Analysis
  8.3.  Misreporting ECN Markings
9.  References
  9.1.  Normative References
  9.2.  Informative References
Appendix A.  Loss Recovery Pseudocode
  A.1.  Tracking Sent Packets
    A.1.1.  Sent Packet Fields
  A.2.  Constants of Interest
  A.3.  Variables of Interest
  A.4.  Initialization
  A.5.  On Sending a Packet
  A.6.  On Receiving a Datagram
  A.7.  On Receiving an Acknowledgment
  A.8.  Setting the Loss Detection Timer
  A.9.  On Timeout
  A.10. Detecting Lost Packets
  A.11. Upon Dropping Initial or Handshake Keys
Appendix B.  Congestion Control Pseudocode
  B.1.  Constants of Interest
  B.2.  Variables of Interest
  B.3.  Initialization
  B.4.  On Packet Sent
  B.5.  On Packet Acknowledgment
  B.6.  On New Congestion Event
  B.7.  Process ECN Information
  B.8.  On Packets Lost
  B.9.  Removing Discarded Packets from Bytes in Flight
Contributors
Authors' Addresses

1. Introduction 🌸

Okay so, QUIC is a secure, general-purpose transport protocol, and it's described in [QUIC-TRANSPORT]. This document is the part that explains its loss detection and congestion control mechanisms. Iconic, honestly 💖

2. Conventions and Definitions 📖

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here. So the caps are doing real work here, bestie 🙈

Here are the terms used in this document:

Ack-eliciting frames: All frames other than ACK, PADDING, and CONNECTION_CLOSE are considered ack-eliciting. So basically everything except those three 💅
Ack-eliciting packets: Packets that contain ack-eliciting frames elicit an ACK from the receiver within the maximum acknowledgment delay and are called ack-eliciting packets. 🔥
In-flight packets: Packets are considered in flight when they are ack-eliciting or contain a PADDING frame, and they have been sent but are not acknowledged, declared lost, or discarded along with old keys. ✈️

3. Design of the QUIC Transmission Machinery 🔧

Every transmission in QUIC is sent with a packet-level header. That header tells you the encryption level and includes a packet sequence number (we'll call it a packet number below). The encryption level tells you which packet number space you're in, as described in Section 12.3 of [QUIC-TRANSPORT]. Packet numbers never repeat within a packet number space for the lifetime of a connection. They're sent in monotonically increasing order within a space, which prevents ambiguity. It's permitted for some packet numbers to never be used, which leaves intentional gaps 🦋

This design means there's no need to tell apart transmissions and retransmissions. That gets rid of a ton of complexity from QUIC's take on TCP loss detection mechanisms, so it's giving clean 🧚

QUIC packets can contain multiple frames of different types. The recovery mechanisms make sure that data and frames needing reliable delivery get acknowledged or declared lost, and then sent in new packets as necessary. The types of frames in a packet affect recovery and congestion control logic, so here's the tea:

All packets are acknowledged, though packets that contain no ack-eliciting frames are only acknowledged along with ack-eliciting packets. 💌
Long header packets that contain CRYPTO frames are critical to the performance of the QUIC handshake and use shorter timers for acknowledgment. They're the main character of the handshake 👑
Packets containing frames besides ACK or CONNECTION_CLOSE frames count toward congestion control limits and are considered to be in flight. 🚦
PADDING frames cause packets to contribute toward bytes in flight without directly causing an acknowledgment to be sent. 🎀

4. Relevant Differences between QUIC and TCP 🪞

Readers who know TCP's loss detection and congestion control will find algorithms here that parallel well-known TCP ones. Lowkey familiar vibes! However, protocol differences between QUIC and TCP lead to algorithmic differences. Those protocol differences are briefly described below 🌈

4.1. Separate Packet Number Spaces

QUIC uses separate packet number spaces for each encryption level, except 0-RTT and all generations of 1-RTT keys use the same packet number space. Separate packet number spaces ensure that acknowledging packets sent with one level of encryption won't cause spurious retransmission of packets sent with a different encryption level. Congestion control and round-trip time (RTT) measurement are unified across packet number spaces. Iykyk, it's all one team 🫶

4.2. Monotonically Increasing Packet Numbers

TCP mixes up transmission order at the sender with delivery order at the receiver. That causes the retransmission ambiguity problem [RETRANSMISSION]. QUIC splits them apart: packet numbers show transmission order, and delivery order is determined by the stream offsets in STREAM frames. Separated and ready 💅

QUIC's packet number is strictly increasing within a packet number space and directly encodes transmission order. A higher packet number means the packet was sent later, and a lower packet number means it was sent earlier. When a packet containing ack-eliciting frames is detected lost, QUIC includes the necessary frames in a new packet with a new packet number. That removes the ambiguity about which packet gets acknowledged when an ACK comes in. Because of this, more accurate RTT measurements can be made, spurious retransmissions are trivially detected, and mechanisms such as Fast Retransmit can be applied universally, based only on packet number. Literally so clean 😭

This design point significantly simplifies loss detection mechanisms for QUIC. Most TCP mechanisms implicitly try to infer transmission ordering from TCP sequence numbers, which is a nontrivial task, especially when TCP timestamps are not available. Big mood for QUIC 🍓

4.3. Clearer Loss Epoch

QUIC starts a loss epoch when a packet is lost. The loss epoch ends when any packet sent after the start of the epoch is acknowledged. TCP waits for the gap in the sequence number space to be filled, so if a segment is lost multiple times in a row, the loss epoch may not end for several round trips. Both should reduce their congestion windows only once per epoch. QUIC will do it once for every round trip that experiences loss, while TCP may only do it once across multiple round trips. QUIC is giving organized 📅

4.4. No Reneging

QUIC ACK frames contain information similar to that in TCP Selective Acknowledgments (SACKs) [RFC2018]. The difference is that QUIC doesn't let a packet acknowledgment be reneged. That makes implementations on both sides way simpler and cuts down memory pressure on the sender 🧠✨

4.5. More ACK Ranges

QUIC supports many ACK ranges, unlike TCP's three SACK ranges. In high-loss environments, this speeds up recovery, reduces spurious retransmits, and ensures forward progress without relying on timeouts. Ranges for days 🌟

4.6. Explicit Correction for Delayed Acknowledgments

QUIC endpoints measure the delay between when a packet is received and when the corresponding acknowledgment is sent. That lets a peer keep a more accurate RTT estimate. See Section 13.2 of [QUIC-TRANSPORT] for the details 🦋

okay bestie, we're picking up right where the last part left off 💅✨

4.7. Probe Timeout Replaces RTO and TLP 🦋

okay so QUIC uses a probe timeout (PTO) 🔥 (see Section 6.2) 💖. The timer is based on how TCP works out its retransmission timeout (RTO) 🎀, which you can read about in [RFC6298]. the PTO also includes the peer's maximum expected acknowledgment delay, so it doesn't just use some fixed minimum timeout like TCP might. no cap, that's giving precision 😭

Now here's the thing: QUIC works kind of like the RACK-TLP loss detection algorithm for TCP [RFC8985] 🌸, and it does NOT collapse the congestion window when the PTO expires 🙅‍♀️. why? because one packet lost at the tail doesn't mean there's persistent congestion, obviously 🙄. QUIC only collapses the congestion window when persistent congestion is declared, see Section 7.6 💅. this avoids unnecessary congestion window reductions, so QUIC doesn't need correcting mechanisms like Forward RTO-Recovery (F-RTO) [RFC5682] 🧚. iconic, honestly.

Since QUIC doesn't collapse the congestion window on a PTO expiration, a QUIC sender is not limited from sending more in-flight packets after a PTO expiration if it still has available congestion window 🫶. this happens when a sender is application limited and the PTO timer expires. ngl this is more aggressive than TCP's RTO mechanism when application limited 🔥, but it's identical when it's not application limited ✨.

QUIC lets probe packets temporarily exceed the congestion window whenever the timer expires 🌈.

4.8. The Minimum Congestion Window Is Two Packets 🍓

TCP uses a minimum congestion window of one packet 💀. but here's the problem: if that single packet gets lost, the sender has to wait for a PTO to recover (Section 6.2) 😭, and that can be way longer than an RTT. also, sending a single ack-eliciting packet makes it more likely you'll get extra latency when a receiver delays its acknowledgment 🙈.

so QUIC recommends that the minimum congestion window be two packets 👑. yes it adds network load, but it's considered safe because the sender still reduces its sending rate exponentially under persistent congestion (Section 6.2) 💜.

4.9. Handshake Packets Are Not Special 💅

TCP treats the loss of a SYN or SYN-ACK packet as persistent congestion and drops the congestion window to one packet; see [RFC5681] 😵. QUIC treats loss of a packet containing handshake data the same as any other loss 🫶. no special treatment, no drama, periodt 💖.

5. Estimating the Round-Trip Time ⭐

okay so at a high level, an endpoint measures the time from when a packet was sent to when it gets acknowledged 🎀. that's an RTT sample ✨. the endpoint uses RTT samples and peer-reported host delays (see Section 13.2 of [QUIC-TRANSPORT]) to build a statistical picture of the network path's RTT 📊. for each path, an endpoint computes three values: the minimum value over a period of time (min_rtt) 🌸, an exponentially weighted moving average (smoothed_rtt) 🦋, and the mean deviation (called "variation" for the rest of this document) in the observed RTT samples (rttvar) 💜. main character energy, fr fr.

5.1. Generating RTT Samples 🧚

an endpoint generates an RTT sample when it receives an ACK frame that meets both of these conditions 👑:

the largest acknowledged packet number is newly acknowledged, and
at least one of the newly acknowledged packets was ack-eliciting.

the RTT sample, latest_rtt, is the time elapsed since the largest acknowledged packet was sent 🔥:

latest_rtt = ack_time - send_time_of_largest_acked

an RTT sample is generated using only the largest acknowledged packet in the received ACK frame 🍓. that's because a peer only reports acknowledgment delays for the largest acknowledged packet in an ACK frame 🙄. the reported acknowledgment delay isn't used by the RTT sample measurement itself, but it does get used to adjust the RTT sample in later smoothed_rtt and rttvar computations (Section 5.3) ✨.

to avoid generating multiple RTT samples for a single packet, an ACK frame SHOULD NOT be used to update RTT estimates if it doesn't newly acknowledge the largest acknowledged packet 🙅‍♀️.

an RTT sample MUST NOT be generated on receiving an ACK frame that doesn't newly acknowledge at least one ack-eliciting packet 🛑. a peer usually doesn't send an ACK frame when only non-ack-eliciting packets are received 🙈. so an ACK frame with acknowledgments for only non-ack-eliciting packets could include an arbitrarily large ACK Delay value 😭. ignoring those ACK frames avoids complications in later smoothed_rtt and rttvar computations 💅.

a sender might generate multiple RTT samples per RTT when multiple ACK frames arrive within an RTT 🌈. as suggested in [RFC6298], doing so might leave smoothed_rtt and rttvar with inadequate history 🥲. keeping RTT estimates with sufficient history is an open research question, so we're not all the way there yet 🧠.

5.2. Estimating min_rtt 🌸

min_rtt is the sender's estimate of the minimum RTT observed for a given network path over a period of time 🎀. in this document, loss detection uses min_rtt to reject implausibly small RTT samples 🫶.

min_rtt MUST be set to the latest_rtt on the first RTT sample 💖. on all other samples, min_rtt MUST be set to the lesser of min_rtt and latest_rtt (Section 5.1) 👑.

an endpoint only uses locally observed times when computing min_rtt, and it does NOT adjust for acknowledgment delays reported by the peer 🙅‍♀️. this lets the endpoint set a lower bound for smoothed_rtt based purely on what it observes (see Section 5.3), and limits potential underestimation from erroneously reported delays by the peer 🦋.

a path's RTT can change over time 🌊. if the actual RTT decreases, min_rtt adapts immediately on the first low sample ✨. if the actual RTT increases, though, min_rtt won't adapt to it 😤, which means future RTT samples that are smaller than the new RTT can still get included in smoothed_rtt 🙃.

Endpoints SHOULD set min_rtt to the newest RTT sample after persistent congestion is established 🔥. this stops the endpoint from repeatedly declaring persistent congestion when the RTT goes up 🙌. it also lets a connection reset its estimates of min_rtt and smoothed_rtt after a disruptive network event; see Section 5.3 💅.

Endpoints MAY reestablish min_rtt at other times in the connection 🌸, like when traffic volume is low and an acknowledgment comes back with a low acknowledgment delay 🍓. implementations SHOULD NOT refresh the min_rtt value too often 🙈, since the actual minimum RTT of the path isn't frequently observable 🥹.

5.3. Estimating smoothed_rtt and rttvar 🦋

smoothed_rtt is an exponentially weighted moving average of an endpoint's RTT samples 💜, and rttvar estimates the variation in those RTT samples using a mean variation 🎀.

the smoothed_rtt calculation uses RTT samples after they've been adjusted for acknowledgment delays ✨. these delays are decoded from the ACK Delay field of ACK frames, as described in Section 19.3 of [QUIC-TRANSPORT] 🧚.

the peer might report acknowledgment delays that are larger than the peer's max_ack_delay during the handshake (Section 13.2.1 of [QUIC-TRANSPORT]) 😳. to account for this, the endpoint SHOULD ignore max_ack_delay until the handshake is confirmed, as defined in Section 4.1.2 of [QUIC-TLS] 🫣. when these big acknowledgment delays happen, they're likely to be non-repeating and limited to the handshake 🙏. so the endpoint can use them without capping them at max_ack_delay, which avoids unnecessarily inflating the RTT estimate 🌈.

heads up: a large acknowledgment delay can seriously inflate smoothed_rtt if there's an error either in the peer's reporting of the delay or in the endpoint's min_rtt estimate 🚨. so before handshake confirmation, an endpoint MAY ignore RTT samples if adjusting them for acknowledgment delay makes the sample smaller than min_rtt 🙅‍♀️.

after the handshake is confirmed, any acknowledgment delays the peer reports that are greater than the peer's max_ack_delay get attributed to unintentional but potentially repeating delays 😤, like scheduler latency at the peer or loss of previous acknowledgments 💀. excess delays could also come from a noncompliant receiver 👀. so these extra delays are treated as effectively part of the path delay and folded into the RTT estimate 🧠.

so, when adjusting an RTT sample using peer-reported acknowledgment delays, an endpoint:

MAY ignore the acknowledgment delay for Initial packets, since these acknowledgments aren't delayed by the peer (Section 13.2.1 of [QUIC-TRANSPORT]) 🍼;
SHOULD ignore the peer's max_ack_delay until the handshake is confirmed;
MUST use the lesser of the acknowledgment delay and the peer's max_ack_delay after the handshake is confirmed; and
MUST NOT subtract the acknowledgment delay from the RTT sample if the result is smaller than min_rtt 🛑. this limits how much smoothed_rtt can be underestimated by a misreporting peer 🫶.

also, an endpoint might postpone processing acknowledgments when the matching decryption keys aren't available yet 🔐. for example, a client might get an acknowledgment for a 0-RTT packet it can't decrypt because its 1-RTT packet protection keys aren't ready yet 🥺. in cases like that, an endpoint SHOULD subtract those local delays from its RTT sample until the handshake is confirmed 💖.

similar to [RFC6298], smoothed_rtt and rttvar get computed like this 🧮.

an endpoint initializes the RTT estimator during connection establishment, and again when the estimator gets reset during connection migration; see Section 9.4 of [QUIC-TRANSPORT] 🦋. before any RTT samples are available for a new path, or when the estimator is reset, it gets initialized using the initial RTT; see Section 6.2.2 🌸.

smoothed_rtt and rttvar start out like this, where kInitialRtt holds the initial RTT value 🎀:

smoothed_rtt = kInitialRtt
rttvar = kInitialRtt / 2

RTT samples for the network path are recorded in latest_rtt; see Section 5.1 📝. on the first RTT sample after initialization, the estimator gets reset using that sample ✨, so it keeps zero history of past samples 🧹. packets sent on other paths don't contribute RTT samples to the current path, as described in Section 9.4 of [QUIC-TRANSPORT] 🙅‍♀️.

on the first RTT sample after initialization, smoothed_rtt and rttvar get set like this 🫶:

smoothed_rtt = latest_rtt
rttvar = latest_rtt / 2

on subsequent RTT samples, smoothed_rtt and rttvar evolve like this 🔥:

ack_delay = decoded acknowledgment delay from ACK frame
if (handshake confirmed):
  ack_delay = min(ack_delay, max_ack_delay)
adjusted_rtt = latest_rtt
if (latest_rtt >= min_rtt + ack_delay):
  adjusted_rtt = latest_rtt - ack_delay
smoothed_rtt = 7/8 * smoothed_rtt + 1/8 * adjusted_rtt
rttvar_sample = abs(smoothed_rtt - adjusted_rtt)
rttvar = 3/4 * rttvar + 1/4 * rttvar_sample

6. Loss Detection 🕵️‍♀️

okay so QUIC senders use acknowledgments to detect lost packets 🔍, and a PTO to make sure acknowledgments actually get received 💌; see Section 6.2. this section describes those algorithms 📖.

if a packet is lost, the QUIC transport needs to recover from that loss 💪. that could mean retransmitting the data, sending an updated frame, or just discarding the frame 🗑️. for more info, check Section 13.3 of [QUIC-TRANSPORT] 🫶.

loss detection is kept separate per packet number space, unlike RTT measurement and congestion control 🧩. the reason is that RTT and congestion control are properties of the path, whereas loss detection also depends on key availability 🔐✨.

6.1. Acknowledgment-Based Detection ✨

okay so acknowledgment-based loss detection is giving TCP's Fast Retransmit [RFC5681], Early Retransmit [RFC5827], Forward Acknowledgment [FACK], SACK loss recovery [RFC6675], and RACK-TLP [RFC8985] energy. It's the same spirit, just adapted for QUIC. here's how it works 💅

A packet is declared lost if it meets all of the following conditions:

The packet is unacknowledged, in flight, and was sent prior to an acknowledged packet.
The packet was sent kPacketThreshold packets before an acknowledged packet (Section 6.1.1), or it was sent long enough in the past (Section 6.1.2).

The acknowledgment tells you that a packet sent later got delivered 🦋 The packet and time thresholds give some tolerance for packet reordering, so the sender doesn't panic over every little shuffle.

Spuriously declaring packets lost leads to unnecessary retransmissions, and it may tank performance because the congestion controller reacts to that loss. Implementations can detect spurious retransmissions and increase the packet or time reordering threshold, so it happens less in the future. Implementations with adaptive time thresholds MAY choose to start with smaller initial reordering thresholds to minimize recovery latency. ⚡

6.1.1. Packet Threshold 🎀

The RECOMMENDED initial value for the packet reordering threshold (kPacketThreshold) is 3, based on best practices for TCP loss detection [RFC5681] [RFC6675]. To stay similar to TCP, implementations SHOULD NOT use a packet threshold less than 3; see [RFC5681].

Some networks have higher degrees of packet reordering, so a sender might detect spurious losses 🙈 Reordering could also be more common with QUIC than TCP. That's because network elements that could observe and reorder TCP packets can't do that for QUIC, and because QUIC packet numbers are encrypted. Algorithms that increase the reordering threshold after spuriously detecting losses, like RACK [RFC8985], have proven useful in TCP and are expected to be at least as useful in QUIC 🌸

6.1.2. Time Threshold ⏰

Once a later packet within the same packet number space has been acknowledged, an endpoint SHOULD declare an earlier packet lost if it was sent a threshold amount of time in the past. To avoid declaring packets lost too early, this time threshold MUST be set to at least the local timer granularity, as indicated by the kGranularity constant. The time threshold is:

max(kTimeThreshold * max(smoothed_rtt, latest_rtt), kGranularity)

If packets sent prior to the largest acknowledged packet cannot yet be declared lost, then a timer SHOULD be set for the remaining time.

Using max(smoothed_rtt, latest_rtt) protects from the two following cases:

the latest RTT sample is lower than the smoothed RTT, perhaps due to reordering where the acknowledgment encountered a shorter path;
the latest RTT sample is higher than the smoothed RTT, perhaps due to a sustained increase in the actual RTT, but the smoothed RTT has not yet caught up.

The RECOMMENDED time threshold (kTimeThreshold), expressed as an RTT multiplier, is 9/8. The RECOMMENDED value of the timer granularity (kGranularity) is 1 millisecond. 🍓

Note: TCP's RACK [RFC8985] specifies a slightly larger threshold, equivalent to 5/4, for a similar purpose. Experience with QUIC shows that 9/8 works well.

Implementations MAY experiment with absolute thresholds, thresholds from previous connections, adaptive thresholds, or the including of RTT variation. Smaller thresholds reduce reordering resilience and increase spurious retransmissions, while larger thresholds increase loss detection delay. It's a tradeoff, bestie ⚖️

6.2. Probe Timeout 🔍

A Probe Timeout (PTO) triggers the sending of one or two probe datagrams when ack-eliciting packets are not acknowledged within the expected period of time, or when the server may not have validated the client's address. A PTO lets a connection recover from loss of tail packets or acknowledgments. it's the comeback arc 🧚

As with loss detection, the PTO is per packet number space. That is, a PTO value is computed per packet number space.

A PTO timer expiration event does not indicate packet loss and MUST NOT cause prior unacknowledged packets to be marked as lost. When an acknowledgment newly acknowledges packets, loss detection proceeds as dictated by the packet and time threshold mechanisms; see Section 6.1.

The PTO algorithm used in QUIC covers the reliability jobs of Tail Loss Probe [RFC8985], RTO [RFC5681], and F-RTO algorithms for TCP [RFC5682]. The timeout computation is based on TCP's RTO period [RFC6298]. 💜

6.2.1. Computing PTO 🧮

When an ack-eliciting packet is transmitted, the sender schedules a timer for the PTO period as follows:

PTO = smoothed_rtt + max(4*rttvar, kGranularity) + max_ack_delay

The PTO period is the amount of time that a sender ought to wait for an acknowledgment of a sent packet. This time period includes the estimated network RTT (smoothed_rtt), the variation in the estimate (4*rttvar), and max_ack_delay, which accounts for the maximum time by which a receiver might delay sending an acknowledgment. ⏳

When the PTO is armed for Initial or Handshake packet number spaces, the max_ack_delay in the PTO period computation is set to 0, since the peer is expected to not delay these packets intentionally; see Section 13.2.1 of [QUIC-TRANSPORT].

The PTO period MUST be at least kGranularity to avoid the timer expiring immediately.

When ack-eliciting packets in multiple packet number spaces are in flight, the timer MUST be set to the earlier value of the Initial and Handshake packet number spaces.

An endpoint MUST NOT set its PTO timer for the Application Data packet number space until the handshake is confirmed. Doing so prevents the endpoint from retransmitting information in packets when either the peer does not yet have the keys to process them or the endpoint does not yet have the keys to process their acknowledgments. For example, this can happen when a client sends 0-RTT packets to the server; it does so without knowing whether the server will be able to decrypt them. Similarly, this can happen when a server sends 1-RTT packets before confirming that the client has verified the server's certificate and can therefore read these 1-RTT packets. the keys have to be ready first, no cap 🔐

A sender SHOULD restart its PTO timer every time an ack-eliciting packet is sent or acknowledged, or when Initial or Handshake keys are discarded (Section 4.9 of [QUIC-TLS]). This ensures the PTO is always set based on the latest estimate of the RTT and for the correct packet across packet number spaces. ✨

When a PTO timer expires, the PTO backoff MUST be increased, resulting in the PTO period being set to twice its current value. The PTO backoff factor is reset when an acknowledgment is received, except in the following case. A server might take longer to respond to packets during the handshake than otherwise. To protect such a server from repeated client probes, the PTO backoff is not reset at a client that is not yet certain that the server has finished validating the client's address. That is, a client does not reset the PTO backoff factor on receiving acknowledgments in Initial packets. 🫶

This exponential reduction in the sender's rate is important because consecutive PTOs might be caused by loss of packets or acknowledgments due to severe congestion. Even when there are ack-eliciting packets in flight in multiple packet number spaces, the exponential increase in PTO occurs across all spaces to prevent excess load on the network. For example, a timeout in the Initial packet number space doubles the length of the timeout in the Handshake packet number space. not spamming the network, periodt 🚫

The total length of time over which consecutive PTOs expire is limited by the idle timeout.

The PTO timer MUST NOT be set if a timer is set for time threshold loss detection; see Section 6.1.2. A timer that is set for time threshold loss detection will expire earlier than the PTO timer in most cases and is less likely to spuriously retransmit data. 👑

6.2.2. Handshakes and New Paths 🤝

Resumed connections over the same network MAY use the previous connection's final smoothed RTT value as the resumed connection's initial RTT. When no previous RTT is available, the initial RTT SHOULD be set to 333 milliseconds. This results in handshakes starting with a PTO of 1 second, as recommended for TCP's initial RTO; see Section 2 of [RFC6298]. 🌈

A connection MAY use the delay between sending a PATH_CHALLENGE and receiving a PATH_RESPONSE to set the initial RTT (see kInitialRtt in Appendix A.2) for a new path, but the delay SHOULD NOT be considered an RTT sample.

When the Initial keys and Handshake keys are discarded (see Section 6.4), any Initial packets and Handshake packets can no longer be acknowledged, so they are removed from bytes in flight. When Initial or Handshake keys are discarded, the PTO and loss detection timers MUST be reset, because discarding keys indicates forward progress and the loss detection timer might have been set for a now-discarded packet number space. moving on means resetting the timers, obviously 🧹

6.2.2.1. Before Address Validation 🛡️

Until the server has validated the client's address on the path, the amount of data it can send is limited to three times the amount of data received, as specified in Section 8.1 of [QUIC-TRANSPORT]. If no additional data can be sent, the server's PTO timer MUST NOT be armed until datagrams have been received from the client, because packets sent on PTO count against the anti-amplification limit. 🚧

When the server receives a datagram from the client, the amplification limit is increased and the server resets the PTO timer. If the PTO timer is then set to a time in the past, it is executed immediately. Doing so avoids sending new 1-RTT packets prior to packets critical to the completion of the handshake. In particular, this can happen when 0-RTT is accepted but the server fails to validate the client's address. 🎀

Since the server could be blocked until more datagrams are received from the client, it is the client's responsibility to send packets to unblock the server until it is certain that the server has finished its address validation (see Section 8 of [QUIC-TRANSPORT]). That is, the client MUST set the PTO timer if the client has not received an acknowledgment for any of its Handshake packets and the handshake is not confirmed (see Section 4.1.2 of [QUIC-TLS]), even if there are no packets in flight. When the PTO fires, the client MUST send a Handshake packet if it has Handshake keys, otherwise it MUST send an Initial packet in a UDP datagram with a payload of at least 1200 bytes. the client is literally carrying the handshake through, that's main character energy 🦋💅

6.2.3. Speeding up Handshake Completion ✨

okay so bestie, picture this: a server gets an Initial packet that has CRYPTO data it already saw 🙈 that's a sign the client didn't get all of the server's CRYPTO data sent in Initial packets, OR the client's estimated RTT is way too small 💀 same vibe on the other side: when a client gets Handshake or 1-RTT packets before it has Handshake keys, it may assume some or all of the server's Initial packets got lost 😭

so to speed up the handshake under these conditions, an endpoint MAY, for a limited number of times per connection, send a packet with unacknowledged CRYPTO data earlier than the PTO expiry 🏃‍♀️💨 but it's still subject to the address validation limits in Section 8.1 of [QUIC-TRANSPORT] 🔒 doing it at most once per connection is adequate to recover quickly from a single packet loss ✨ and an endpoint that always retransmits in response to packets it can't process risks creating an infinite exchange of packets 😱 literally a never-ending ping-pong, not it 🙅‍♀️

endpoints can also use coalesced packets (see Section 12.2 of [QUIC-TRANSPORT]) so that each datagram elicits at least one acknowledgment 🫶 for example, a client can coalesce an Initial packet with a PING and PADDING frames together with a 0-RTT data packet 🌸 and a server can coalesce an Initial packet with a PING frame alongside one or more packets in its first flight 🧚

6.2.4. Sending Probe Packets 💌

okay this part is very important bestie 💅 when a PTO timer expires, a sender MUST send at least one ack-eliciting packet in the packet number space as a probe 🔥 an endpoint MAY send up to two full-sized datagrams containing ack-eliciting packets, so it avoids an expensive consecutive PTO expiration because of a single lost datagram, or to send data from multiple packet number spaces 🎀 all probe packets sent on a PTO MUST be ack-eliciting 💜

in addition to sending data in the packet number space where the timer expired, the sender SHOULD send ack-eliciting packets from other packet number spaces that have in-flight data, coalescing packets if it can 🦋 this is especially valuable when the server has both Initial and Handshake data in flight, or when the client has both Handshake and Application Data in flight, because the peer might only have receive keys for one of those two packet number spaces 🙈

if the sender wants a faster acknowledgment on PTO, it can skip a packet number to get rid of the acknowledgment delay ⭐

an endpoint SHOULD include new data in packets sent on PTO expiration 💖 previously sent data MAY be sent if no new data can be sent 🍓 implementations MAY use alternative strategies for picking the content of probe packets, including sending new or retransmitted data based on the application's priorities 👑

it's possible the sender has no new or previously sent data to send 🤔 like, imagine this: new application data goes out in a STREAM frame, gets deemed lost, gets retransmitted in a new packet, and then the original transmission gets acknowledged 💀 when there's no data to send, the sender SHOULD send a PING or other ack-eliciting frame in a single packet, which rearms the PTO timer 🔔

alternatively, instead of sending an ack-eliciting packet, the sender MAY mark any packets still in flight as lost 🗑️ this avoids sending an extra packet, but it raises the risk that loss gets declared too aggressively, which leads to an unnecessary rate reduction by the congestion controller 😬

consecutive PTO periods increase exponentially, so connection recovery latency also increases exponentially as packets keep getting dropped in the network 📈 sending two packets on PTO expiration increases resilience to packet drops, which reduces the probability of consecutive PTO events 🛡️

when the PTO timer expires multiple times and no new data can be sent, implementations MUST choose between sending the same payload every time or sending different payloads 🎭 sending the same payload may be simpler and makes sure the highest priority frames arrive first 🥇 sending different payloads each time reduces the chances of spurious retransmission 🌈

6.3. Handling Retry Packets 🔁

a Retry packet makes a client send another Initial packet, which basically restarts the connection process 🔄 a Retry packet tells the client the Initial packet was received but not processed 📬 a Retry packet can't be treated as an acknowledgment, because it doesn't say a packet was processed or specify the packet number 🙅‍♀️

clients that receive a Retry packet reset congestion control and loss recovery state, including any pending timers 🧹 other connection state, especially cryptographic handshake messages, is kept; see Section 17.2.5 of [QUIC-TRANSPORT] 💜

the client MAY compute an RTT estimate to the server as the time from when the first Initial packet was sent to when a Retry or a Version Negotiation packet is received ⏱️ the client MAY use this value in place of its default for the initial RTT estimate 🦋

6.4. Discarding Keys and Packet State 🗝️

when Initial and Handshake packet protection keys get discarded (see Section 4.9 of [QUIC-TLS]), all packets sent with those keys can't be acknowledged anymore, because their acknowledgments can't be processed 💀 the sender MUST discard all recovery state associated with those packets and MUST remove them from the count of bytes in flight 🫧

endpoints stop sending and receiving Initial packets once they start exchanging Handshake packets; see Section 17.2.2.1 of [QUIC-TRANSPORT] 🤝 at this point, recovery state for all in-flight Initial packets is discarded 🧺

when 0-RTT is rejected, recovery state for all in-flight 0-RTT packets is discarded 🙃

if a server accepts 0-RTT but doesn't buffer 0-RTT packets that arrive before Initial packets, early 0-RTT packets will be declared lost 😭 but that's expected to be infrequent 🌷

it's expected that keys get discarded at some point after the packets encrypted with them are either acknowledged or declared lost ✅ however, Initial and Handshake secrets are discarded as soon as Handshake and 1-RTT keys are proven to be available to both client and server; see Section 4.9.1 of [QUIC-TLS] 🎉

7. Congestion Control 🌊

this document specifies a sender-side congestion controller for QUIC that's similar to TCP NewReno [RFC6582] 💖

the signals QUIC gives for congestion control are generic and designed to support different sender-side algorithms 🧩 a sender can unilaterally choose a different algorithm to use, such as CUBIC [RFC8312] 🎀

if a sender uses a different controller than the one specified in this document, the chosen controller MUST conform to the congestion control guidelines specified in Section 3.1 of [RFC8085] 📏

similar to TCP, packets containing only ACK frames don't count toward bytes in flight and aren't congestion controlled 🙌 unlike TCP, QUIC can detect the loss of these packets and MAY use that information to adjust the congestion controller or the rate of ACK-only packets being sent 🔍 but this document doesn't describe a mechanism for doing so 🤷‍♀️

the congestion controller is per path, so packets sent on other paths don't change the current path's congestion controller, as described in Section 9.4 of [QUIC-TRANSPORT] 🛤️

the algorithm in this document specifies and uses the controller's congestion window in bytes 🪣

an endpoint MUST NOT send a packet if it would make bytes_in_flight (see Appendix B.2) larger than the congestion window, unless the packet is sent on a PTO timer expiration (see Section 6.2) or when entering recovery (see Section 7.3.2) 🚫

7.1. Explicit Congestion Notification 🚦

if a path has been validated to support Explicit Congestion Notification (ECN) [RFC3168] [RFC8311], QUIC treats a Congestion Experienced (CE) codepoint in the IP header as a signal of congestion 🔥 this document specifies an endpoint's response when the peer-reported ECN-CE count increases; see Section 13.4.2 of [QUIC-TRANSPORT] 📝

7.2. Initial and Minimum Congestion Window 🪟

QUIC begins every connection in slow start with the congestion window set to an initial value 🚀 endpoints SHOULD use an initial congestion window of ten times the maximum datagram size (max_datagram_size), while limiting the window to the larger of 14,720 bytes or twice the maximum datagram size 💅 this follows the analysis and recommendations in [RFC6928], increasing the byte limit to account for the smaller 8-byte overhead of UDP compared to the 20-byte overhead for TCP 🧠

if the maximum datagram size changes during the connection, the initial congestion window SHOULD be recalculated with the new size 📐 if the maximum datagram size is decreased in order to complete the handshake, the congestion window SHOULD be set to the new initial congestion window 🔄

prior to validating the client's address, the server can be further limited by the anti-amplification limit as specified in Section 8.1 of [QUIC-TRANSPORT] 🛡️ though the anti-amplification limit can stop the congestion window from being fully used and so slow down its growth, it doesn't directly affect the congestion window itself 🐢

the minimum congestion window is the smallest value the congestion window can reach in response to loss, an increase in the peer-reported ECN-CE count, or persistent congestion 📉 the RECOMMENDED value is 2 * max_datagram_size 🎯

7.3. Congestion Control States 🗺️

the NewReno congestion controller described in this document has three distinct states, as shown in Figure 1 🌟

                New path or      +------------+
           persistent congestion |   Slow     |
       (O)---------------------->|   Start    |
                                 +------------+
                                       |
                               Loss or |
                       ECN-CE increase |
                                       v
+------------+     Loss or       +------------+
| Congestion |  ECN-CE increase  |  Recovery  |
| Avoidance  |------------------>|   Period   |
+------------+                   +------------+
          ^                            |
          |                            |
          +----------------------------+
             Acknowledgment of packet
               sent during recovery

        Figure 1: Congestion Control States and Transitions

these states and the transitions between them are described in the subsequent sections 🗒️

7.3.1. Slow Start 🐣

a NewReno sender is in slow start any time the congestion window is below the slow start threshold 🌱 a sender begins in slow start because the slow start threshold is initialized to an infinite value ♾️

while a sender is in slow start, the congestion window increases by the number of bytes acknowledged when each acknowledgment is processed 📈 this results in exponential growth of the congestion window 🚀

the sender MUST exit slow start and enter a recovery period when a packet is lost or when the ECN-CE count reported by its peer increases 🛑

a sender reenters slow start any time the congestion window is less than the slow start threshold, which only happens after persistent congestion is declared 🔁

7.3.2. Recovery ✨

okay so a NewReno sender enters a recovery period when it detects a lost packet 💔 or when the ECN-CE count its peer reports goes up. if it's already in a recovery period, it just chills there and does NOT reenter it. no double-dipping 🙅‍♀️

when a sender enters a recovery period, it MUST set the slow start threshold to half the congestion window value at the moment loss is detected. and before it leaves the recovery period, the congestion window MUST be set to that reduced slow start threshold value. bestie, this part is non-negotiable 💅

implementations MAY slam the congestion window down right away when they enter recovery, or they can use other tools, like Proportional Rate Reduction [PRR], to lower it more gently 🦋 if you reduce it immediately, you're allowed to send one packet before the reduction kicks in. that helps loss recovery go faster if the data in the lost packet gets retransmitted 🚀 this is similar to TCP, described in Section 5 of [RFC6675] 🎀

the whole point of a recovery period is to limit the congestion window reduction to once per round trip 🌸 so during a recovery period, the congestion window doesn't change just because there are new losses or because the ECN-CE count goes up again. it's giving stability 😌

a recovery period ends, and the sender enters congestion avoidance, when a packet that was sent during the recovery period gets acknowledged ✨ ngl this is a little different from TCP, where recovery ends when the lost segment that kicked it off is acknowledged [RFC5681]. not the same vibe, so don't mix them up 🙈

7.3.3. Congestion Avoidance 🌈

a NewReno sender is in congestion avoidance any time its congestion window is at or above the slow start threshold and it is not in a recovery period 💜

a sender in congestion avoidance uses an Additive Increase Multiplicative Decrease (AIMD) approach. it MUST limit the increase to the congestion window to at most one maximum datagram size for each congestion window that gets acknowledged. slow and steady, we love to see it 🐢💖

the sender leaves congestion avoidance and enters a recovery period when a packet is lost or when the ECN-CE count reported by its peer goes up 🚨

7.4. Ignoring Loss of Undecryptable Packets 🙈

okay so during the handshake, some packet protection keys might not be available yet when a packet shows up, and the receiver can choose to drop that packet 🫣 specifically, Handshake and 0-RTT packets can't be processed until the Initial packets arrive, and 1-RTT packets can't be processed until the handshake is done 🔐

endpoints MAY ignore the loss of Handshake, 0-RTT, and 1-RTT packets that might have arrived before the peer had the packet protection keys to process them 🫶 but endpoints MUST NOT ignore the loss of packets that were sent after the earliest acknowledged packet in a given packet number space. that one's a hard no 🙅‍♀️💀

7.5. Probe Timeout 🔔

probe packets MUST NOT be blocked by the congestion controller 🚫 BUT a sender MUST count these packets as being additionally in flight. why? because probes add network load without proving any packet was actually lost, so they still take up room 🧳 also, sending probe packets might push the sender's bytes in flight past the congestion window until an acknowledgment comes back that establishes loss or delivery of packets. it's a whole thing, but it's valid 💅

7.6. Persistent Congestion 🌧️

when a sender establishes loss of all packets sent over a long enough duration, the network is considered to be experiencing persistent congestion. like, the whole network is struggling, not just one packet having a bad day 😭

7.6.1. Duration ⏳

the persistent congestion duration is computed like this:

(smoothed_rtt + max(4*rttvar, kGranularity) + max_ack_delay) *
    kPersistentCongestionThreshold

unlike the PTO computation in Section 6.2, this duration includes the max_ack_delay no matter which packet number spaces the losses were established in 🎀

this duration lets a sender send as many packets as it needs before it decides there's persistent congestion. that includes some packets sent in response to PTO expiration too 🧚 TCP does the same kind of thing with Tail Loss Probes [RFC8985] and an RTO [RFC5681], so we're in good company 🤝

bigger values of kPersistentCongestionThreshold make the sender less responsive to persistent congestion in the network, which can lead to aggressive sending into a congested network 😬 but if the value is too small, the sender might declare persistent congestion unnecessarily, and then its throughput drops for no good reason 📉

the RECOMMENDED value for kPersistentCongestionThreshold is 3 ✨ that works out to behavior that's roughly the same as a TCP sender declaring an RTO after two TLPs 🦋

this design doesn't use consecutive PTO events to establish persistent congestion, because application patterns affect PTO expiration 🙃 for example, a sender that sends small bits of data with quiet periods in between restarts the PTO timer every time it sends. that could keep the PTO timer from expiring for a really long time, even when no acknowledgments are coming back 😵‍💫 using a duration instead means a sender can establish persistent congestion without depending on PTO expiration. iconic 👑

7.6.2. Establishing Persistent Congestion 🕵️‍♀️

a sender establishes persistent congestion after receiving an acknowledgment if two ack-eliciting packets get declared lost, and:

across all packet number spaces, none of the packets sent between the send times of these two packets are acknowledged;
the duration between the send times of these two packets exceeds the persistent congestion duration (Section 7.6.1); and
a prior RTT sample existed when these two packets were sent.

these two packets MUST be ack-eliciting, since a receiver is required to acknowledge only ack-eliciting packets within its maximum acknowledgment delay; see Section 13.2 of [QUIC-TRANSPORT] 🫡

the persistent congestion period SHOULD NOT start until there's at least one RTT sample. before that first RTT sample, a sender arms its PTO timer based on the initial RTT (Section 6.2.2), and that could be a lot bigger than the actual RTT 😅 requiring a prior RTT sample keeps a sender from declaring persistent congestion with maybe too few probes. not on my watch 🙅‍♀️

since network congestion doesn't care about packet number spaces, persistent congestion SHOULD consider packets sent across packet number spaces 🌐 a sender that doesn't have state for all packet number spaces, or an implementation that can't compare send times across packet number spaces, MAY use state for just the packet number space that was acknowledged. this might lead to erroneously declaring persistent congestion, but it won't lead to a failure to detect persistent congestion 🫠 so worst case is it's a little over-cautious, which is kinda fine

when persistent congestion is declared, the sender's congestion window MUST be reduced to the minimum congestion window (kMinimumWindow), similar to how a TCP sender responds to an RTO [RFC5681] 📉💅

7.6.3. Example 📝

here's an example that shows how a sender might establish persistent congestion 🧠 assume:

smoothed_rtt + max(4*rttvar, kGranularity) + max_ack_delay = 2
kPersistentCongestionThreshold = 3

now check out this sequence of events:

           +========+===================================+
           | Time   | Action                            |
           +========+===================================+
           | t=0    | Send packet #1 (application data) |
           +--------+-----------------------------------+
           | t=1    | Send packet #2 (application data) |
           +--------+-----------------------------------+
           | t=1.2  | Receive acknowledgment of #1      |
           +--------+-----------------------------------+
           | t=2    | Send packet #3 (application data) |
           +--------+-----------------------------------+
           | t=3    | Send packet #4 (application data) |
           +--------+-----------------------------------+
           | t=4    | Send packet #5 (application data) |
           +--------+-----------------------------------+
           | t=5    | Send packet #6 (application data) |
           +--------+-----------------------------------+
           | t=6    | Send packet #7 (application data) |
           +--------+-----------------------------------+
           | t=8    | Send packet #8 (PTO 1)            |
           +--------+-----------------------------------+
           | t=12   | Send packet #9 (PTO 2)            |
           +--------+-----------------------------------+
           | t=12.2 | Receive acknowledgment of #9      |
           +--------+-----------------------------------+

                              Table 1

packets 2 through 8 get declared lost when the acknowledgment for packet 9 arrives at "t = 12.2" 😱

the congestion period is calculated as the time between the oldest and newest lost packets: "8 - 1 = 7". the persistent congestion duration is "2 * 3 = 6". because the threshold was reached, AND none of the packets between the oldest and newest lost packets were acknowledged, the network is considered to have experienced persistent congestion. main character energy 🎬

while this example shows PTO expiration, PTO expiration isn't required for persistent congestion to be established 🌟

7.7. Pacing 🌊

a sender SHOULD pace the sending of all in-flight packets, based on input from the congestion controller 💧

sending a bunch of packets into the network with zero delay between them makes a packet burst, and that can cause short-term congestion and losses 💥 so senders MUST either use pacing or limit those bursts. senders SHOULD limit bursts to the initial congestion window; see Section 7.2. and a sender that knows the network path to the receiver can handle bigger bursts MAY use a higher limit 🫧

an implementation should take care to build its congestion controller so it works well with a pacer 🧁 for instance, a pacer might wrap the congestion controller and control when the congestion window is available, or a pacer might pace out packets that get handed to it by the congestion controller 🍓

timely delivery of ACK frames matters a lot for efficient loss recovery 📬 so, to avoid delaying their delivery to the peer, packets containing only ACK frames SHOULD therefore not be paced 🫶

endpoints can implement pacing however they choose 🦋 a perfectly paced sender spreads packets exactly evenly over time. for a window-based congestion controller, like the one in this document, that rate can be worked out by averaging the congestion window over the RTT 🧮 expressed as a rate in units of bytes per time, where congestion_window is in bytes:

rate = N * congestion_window / smoothed_rtt

or expressed as an inter-packet interval in units of time:

interval = ( smoothed_rtt * packet_size / congestion_window ) / N

using a value for "N" that's small, but at least 1 (for example, 1.25), makes sure that variations in RTT don't result in underutilization of the congestion window 🎯

practical stuff, like packetization, scheduling delays, and computational efficiency, can make a sender deviate from this rate over time periods that are much shorter than an RTT 🙃 it happens, we adapt 💖

one possible implementation strategy for pacing uses a leaky bucket algorithm 🪣 the capacity of the "bucket" is limited to the maximum burst size, and the rate the "bucket" fills is determined by the function above. it's giving organized, it's giving efficient, it's giving everything 🌈✨

7.8. Underutilizing the Congestion Window 🦋

okay so when bytes in flight is smaller than the congestion window AND sending isn't pacing limited, the congestion window is underutilized 🥺 this can happen because the app doesn't have enough data to send, or because of flow control limits 💅. when that happens, the congestion window SHOULD NOT be increased in either slow start or congestion avoidance. no growing the window when you're not even using it, bestie 🙅‍♀️

a sender that paces packets (see Section 7.7) might delay sending packets and not fully use the congestion window because of that delay 🌸. a sender SHOULD NOT consider itself application limited if it would have fully used the congestion window without the pacing delay ✨

a sender MAY implement alternative mechanisms to update its congestion window after periods of underutilization, such as those proposed for TCP in [RFC7661]. the way that TCP does it is a whole other vibe 🧚

8. Security Considerations 🚨

this section is serious, so buckle up 💜

8.1. Loss and Congestion Signals

loss detection and congestion control are built on signals like delay, loss, and ECN markings, and those signals come from entities that aren't authenticated 😬 an attacker can make endpoints slow down their sending rate by manipulating these signals. they can drop packets, change the path delay strategically, or change ECN codepoints. the danger is real, fr fr 🔥

8.2. Traffic Analysis

packets that carry only ACK frames can be heuristically identified by looking at their size 👀 acknowledgment patterns may also reveal information about link characteristics or application behavior. to reduce leaked information, endpoints can bundle acknowledgments with other frames, or they can use PADDING frames, though that comes at a potential cost to performance 🫶

8.3. Misreporting ECN Markings

a receiver can misreport ECN markings to change the congestion response of a sender. suppressing reports of ECN-CE markings could make a sender increase its send rate, and that increase could lead to congestion and loss 💀

a sender can detect suppression of reports by marking some of the packets it sends with an ECN-CE marking. if a packet sent with an ECN-CE marking isn't reported as CE marked when the packet is acknowledged, the sender can disable ECN for that path by not setting ECN-Capable Transport (ECT) codepoints in later packets sent on that path [RFC3168]. detective energy 🕵️‍♀️

reporting additional ECN-CE markings makes a sender reduce its sending rate. that works a lot like advertising reduced connection flow control limits, so there's no advantage to doing it 🙅‍♀️

endpoints choose the congestion controller they use. congestion controllers respond to reports of ECN-CE by reducing their rate, but the response may vary. markings can be treated as equivalent to loss [RFC3168], but other responses can be specified, such as [RFC8511] or [RFC8311]. so many options, bestie 🌈

9. References 📚

9.1. Normative References

[QUIC-TLS] Thomson, M., Ed. and S. Turner, Ed., "Using TLS to Secure QUIC", RFC 9001, DOI 10.17487/RFC9001, May 2021, https://www.rfc-editor.org/info/rfc9001.
[QUIC-TRANSPORT] Iyengar, J., Ed. and M. Thomson, Ed., "QUIC: A UDP-Based Multiplexed and Secure Transport", RFC 9000, DOI 10.17487/RFC9000, May 2021, https://www.rfc-editor.org/info/rfc9000.
[RFC2119] Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", BCP 14, RFC 2119, DOI 10.17487/RFC2119, March 1997, https://www.rfc-editor.org/info/rfc2119.
[RFC3168] Ramakrishnan, K., Floyd, S., and D. Black, "The Addition of Explicit Congestion Notification (ECN) to IP", RFC 3168, DOI 10.17487/RFC3168, September 2001, https://www.rfc-editor.org/info/rfc3168.
[RFC8085] Eggert, L., Fairhurst, G., and G. Shepherd, "UDP Usage Guidelines", BCP 145, RFC 8085, DOI 10.17487/RFC8085, March 2017, https://www.rfc-editor.org/info/rfc8085.
[RFC8174] Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC 2119 Key Words", BCP 14, RFC 8174, DOI 10.17487/RFC8174, May 2017, https://www.rfc-editor.org/info/rfc8174.

9.2. Informative References

[FACK] Mathis, M. and J. Mahdavi, "Forward acknowledgement: Refining TCP Congestion Control", ACM SIGCOMM Computer Communication Review, DOI 10.1145/248157.248181, August 1996, https://doi.org/10.1145/248157.248181.
[PRR] Mathis, M., Dukkipati, N., and Y. Cheng, "Proportional Rate Reduction for TCP", RFC 6937, DOI 10.17487/RFC6937, May 2013, https://www.rfc-editor.org/info/rfc6937.
[RETRANSMISSION] Karn, P. and C. Partridge, "Improving Round-Trip Time Estimates in Reliable Transport Protocols", ACM Transactions on Computer Systems, DOI 10.1145/118544.118549, November 1991, https://doi.org/10.1145/118544.118549.
[RFC2018] Mathis, M., Mahdavi, J., Floyd, S., and A. Romanow, "TCP Selective Acknowledgment Options", RFC 2018, DOI 10.17487/RFC2018, October 1996, https://www.rfc-editor.org/info/rfc2018.
[RFC3465] Allman, M., "TCP Congestion Control with Appropriate Byte Counting (ABC)", RFC 3465, DOI 10.17487/RFC3465, February 2003, https://www.rfc-editor.org/info/rfc3465.
[RFC5681] Allman, M., Paxson, V., and E. Blanton, "TCP Congestion Control", RFC 5681, DOI 10.17487/RFC5681, September 2009, https://www.rfc-editor.org/info/rfc5681.
[RFC5682] Sarolahti, P., Kojo, M., Yamamoto, K., and M. Hata, "Forward RTO-Recovery (F-RTO): An Algorithm for Detecting Spurious Retransmission Timeouts with TCP", RFC 5682, DOI 10.17487/RFC5682, September 2009, https://www.rfc-editor.org/info/rfc5682.
[RFC5827] Allman, M., Avrachenkov, K., Ayesta, U., Blanton, J., and P. Hurtig, "Early Retransmit for TCP and Stream Control Transmission Protocol (SCTP)", RFC 5827, DOI 10.17487/RFC5827, May 2010, https://www.rfc-editor.org/info/rfc5827.
[RFC6298] Paxson, V., Allman, M., Chu, J., and M. Sargent, "Computing TCP's Retransmission Timer", RFC 6298, DOI 10.17487/RFC6298, June 2011, https://www.rfc-editor.org/info/rfc6298.
[RFC6582] Henderson, T., Floyd, S., Gurtov, A., and Y. Nishida, "The NewReno Modification to TCP's Fast Recovery Algorithm", RFC 6582, DOI 10.17487/RFC6582, April 2012, https://www.rfc-editor.org/info/rfc6582.
[RFC6675] Blanton, E., Allman, M., Wang, L., Jarvinen, I., Kojo, M., and Y. Nishida, "A Conservative Loss Recovery Algorithm Based on Selective Acknowledgment (SACK) for TCP", RFC 6675, DOI 10.17487/RFC6675, August 2012, https://www.rfc-editor.org/info/rfc6675.
[RFC6928] Chu, J., Dukkipati, N., Cheng, Y., and M. Mathis, "Increasing TCP's Initial Window", RFC 6928, DOI 10.17487/RFC6928, April 2013, https://www.rfc-editor.org/info/rfc6928.
[RFC7661] Fairhurst, G., Sathiaseelan, A., and R. Secchi, "Updating TCP to Support Rate-Limited Traffic", RFC 7661, DOI 10.17487/RFC7661, October 2015, https://www.rfc-editor.org/info/rfc7661.
[RFC8311] Black, D., "Relaxing Restrictions on Explicit Congestion Notification (ECN) Experimentation", RFC 8311, DOI 10.17487/RFC8311, January 2018, https://www.rfc-editor.org/info/rfc8311.
[RFC8312] Rhee, I., Xu, L., Ha, S., Zimmermann, A., Eggert, L., and R. Scheffenegger, "CUBIC for Fast Long-Distance Networks", RFC 8312, DOI 10.17487/RFC8312, February 2018, https://www.rfc-editor.org/info/rfc8312.
[RFC8511] Khademi, N., Welzl, M., Armitage, G., and G. Fairhurst, "TCP Alternative Backoff with ECN (ABE)", RFC 8511, DOI 10.17487/RFC8511, December 2018, https://www.rfc-editor.org/info/rfc8511.
[RFC8985] Cheng, Y., Cardwell, N., Dukkipati, N., and P. Jha, "The RACK-TLP Loss Detection Algorithm for TCP", RFC 8985, DOI 10.17487/RFC8985, February 2021, https://www.rfc-editor.org/info/rfc8985.

Appendix A. Loss Recovery Pseudocode 🧮

we're about to walk through an example implementation of the loss detection mechanisms described in Section 6 ✨

the pseudocode segments in this section are licensed as Code Components; see the copyright notice 📜

A.1. Tracking Sent Packets

to correctly implement congestion control, a QUIC sender tracks every ack-eliciting packet until that packet is acknowledged or lost 📦 implementations are expected to access this information by packet number and crypto context, and to store the per-packet fields (Appendix A.1.1) for loss recovery and congestion control.

after a packet is declared lost, the endpoint can still keep state for it for a while to allow for packet reordering; see Section 13.3 of [QUIC-TRANSPORT]. this lets a sender detect spurious retransmissions 🕵️‍♀️

sent packets are tracked for each packet number space, and ACK processing only applies to a single space 🌸

A.1.1. Sent Packet Fields

packet_number: the packet number of the sent packet.
ack_eliciting: a Boolean that indicates whether a packet is ack-eliciting. if true, an acknowledgment is expected, though the peer could delay sending the ACK frame containing it by up to the max_ack_delay ⏳
in_flight: a Boolean that indicates whether the packet counts toward bytes in flight.
sent_bytes: the number of bytes sent in the packet, not including UDP or IP overhead, but including QUIC framing overhead.
time_sent: the time the packet was sent ⏰

A.2. Constants of Interest

constants used in loss recovery are based on a combination of RFCs, papers, and common practice 💖

kPacketThreshold: maximum reordering in packets before packet threshold loss detection considers a packet lost. the value recommended in Section 6.1.1 is 3.
kTimeThreshold: maximum reordering in time before time threshold loss detection considers a packet lost. specified as an RTT multiplier. the value recommended in Section 6.1.2 is 9/8.
kGranularity: timer granularity. this is a system-dependent value, and Section 6.1.2 recommends a value of 1 ms.
kInitialRtt: the RTT used before an RTT sample is taken. the value recommended in Section 6.2.2 is 333 ms.
kPacketNumberSpace: an enum to enumerate the three packet number spaces:

enum kPacketNumberSpace {
  Initial,
  Handshake,
  ApplicationData,
}

A.3. Variables of Interest 💜

Okay bestie, these are the variables the congestion control stuff needs to actually run 💅

latest_rtt: The most recent RTT measurement made when receiving an acknowledgment for a previously unacknowledged packet.
smoothed_rtt: The smoothed RTT of the connection, computed as described in Section 5.3.
rttvar: The RTT variation, computed as described in Section 5.3.
min_rtt: The minimum RTT seen over a period of time, ignoring acknowledgment delay, as described in Section 5.2.
first_rtt_sample: The time that the first RTT sample was obtained.
max_ack_delay: The maximum amount of time by which the receiver intends to delay acknowledgments for packets in the Application Data packet number space, as defined by the eponymous transport parameter (Section 18.2 of [QUIC-TRANSPORT]). Note that the actual ack_delay in a received ACK frame may be larger due to late timers, reordering, or loss. 🕰️
loss_detection_timer: Multi-modal timer used for loss detection. ⏰
pto_count: The number of times a PTO has been sent without receiving an acknowledgment.
time_of_last_ack_eliciting_packet[kPacketNumberSpace]: The time the most recent ack-eliciting packet was sent.
largest_acked_packet[kPacketNumberSpace]: The largest packet number acknowledged in the packet number space so far.
loss_time[kPacketNumberSpace]: The time at which the next packet in that packet number space can be considered lost based on exceeding the reordering window in time.
sent_packets[kPacketNumberSpace]: An association of packet numbers in a packet number space to information about them. Described in detail above in Appendix A.1. 🗂️

A.4. Initialization 🌱

At the beginning of the connection, you initialize the loss detection variables like this, no skipping steps 🫶

   loss_detection_timer.reset()
   pto_count = 0
   latest_rtt = 0
   smoothed_rtt = kInitialRtt
   rttvar = kInitialRtt / 2
   min_rtt = 0
   first_rtt_sample = 0
   for pn_space in [ Initial, Handshake, ApplicationData ]:
     largest_acked_packet[pn_space] = infinite
     time_of_last_ack_eliciting_packet[pn_space] = 0
     loss_time[pn_space] = 0

A.5. On Sending a Packet 📤

After a packet is sent, information about the packet gets stored. The parameters to OnPacketSent are described in detail above in Appendix A.1.1. 📝

Pseudocode for OnPacketSent follows:

   OnPacketSent(packet_number, pn_space, ack_eliciting,
                in_flight, sent_bytes):
     sent_packets[pn_space][packet_number].packet_number =
                                              packet_number
     sent_packets[pn_space][packet_number].time_sent = now()
     sent_packets[pn_space][packet_number].ack_eliciting =
                                              ack_eliciting
     sent_packets[pn_space][packet_number].in_flight = in_flight
     sent_packets[pn_space][packet_number].sent_bytes = sent_bytes
     if (in_flight):
       if (ack_eliciting):
         time_of_last_ack_eliciting_packet[pn_space] = now()
       OnPacketSentCC(sent_bytes)
       SetLossDetectionTimer()

A.6. On Receiving a Datagram 📨

Okay, so this one is giving rescue energy 🦸‍♀️ When a server is blocked by anti-amplification limits, receiving a datagram unblocks it, even if none of the packets in that datagram get successfully processed. In that case, the PTO timer needs to be rearmed. Iconic. 🔥

Pseudocode for OnDatagramReceived follows:

   OnDatagramReceived(datagram):
     // If this datagram unblocks the server, arm the
     // PTO timer to avoid deadlock.
     if (server was at anti-amplification limit):
       SetLossDetectionTimer()
       if loss_detection_timer.timeout < now():
         // Execute PTO if it would have expired
         // while the amplification limit applied.
         OnLossDetectionTimeout()

A.7. On Receiving an Acknowledgment 🧾

When an ACK frame is received, it may newly acknowledge any number of packets. Lots of packets, lots of drama 🎭

Pseudocode for OnAckReceived and UpdateRtt follow:

   IncludesAckEliciting(packets):
     for packet in packets:
       if (packet.ack_eliciting):
         return true
     return false

   OnAckReceived(ack, pn_space):
     if (largest_acked_packet[pn_space] == infinite):
       largest_acked_packet[pn_space] = ack.largest_acked
     else:
       largest_acked_packet[pn_space] =
           max(largest_acked_packet[pn_space], ack.largest_acked)

     // DetectAndRemoveAckedPackets finds packets that are newly
     // acknowledged and removes them from sent_packets.
     newly_acked_packets =
         DetectAndRemoveAckedPackets(ack, pn_space)
     // Nothing to do if there are no newly acked packets.
     if (newly_acked_packets.empty()):
       return

     // Update the RTT if the largest acknowledged is newly acked
     // and at least one ack-eliciting was newly acked.
     if (newly_acked_packets.largest().packet_number ==
             ack.largest_acked &&
         IncludesAckEliciting(newly_acked_packets)):
       latest_rtt =
         now() - newly_acked_packets.largest().time_sent
       UpdateRtt(ack.ack_delay)

     // Process ECN information if present.
     if (ACK frame contains ECN information):
         ProcessECN(ack, pn_space)

     lost_packets = DetectAndRemoveLostPackets(pn_space)
     if (!lost_packets.empty()):
       OnPacketsLost(lost_packets)
     OnPacketsAcked(newly_acked_packets)

     // Reset pto_count unless the client is unsure if
     // the server has validated the client's address.
     if (PeerCompletedAddressValidation()):
       pto_count = 0
     SetLossDetectionTimer()


   UpdateRtt(ack_delay):
     if (first_rtt_sample == 0):
       min_rtt = latest_rtt
       smoothed_rtt = latest_rtt
       rttvar = latest_rtt / 2
       first_rtt_sample = now()
       return

     // min_rtt ignores acknowledgment delay.
     min_rtt = min(min_rtt, latest_rtt)
     // Limit ack_delay by max_ack_delay after handshake
     // confirmation.
     if (handshake confirmed):
       ack_delay = min(ack_delay, max_ack_delay)

     // Adjust for acknowledgment delay if plausible.
     adjusted_rtt = latest_rtt
     if (latest_rtt >= min_rtt + ack_delay):
       adjusted_rtt = latest_rtt - ack_delay

     rttvar = 3/4 * rttvar + 1/4 * abs(smoothed_rtt - adjusted_rtt)
     smoothed_rtt = 7/8 * smoothed_rtt + 1/8 * adjusted_rtt

A.8. Setting the Loss Detection Timer ⏲️

QUIC loss detection uses a single timer for all timeout loss detection. This one timer is the MVP 🏆 Its duration depends on the timer's mode, which gets set in the packet and timer events further below. The function SetLossDetectionTimer, defined below, shows how that single timer gets set.

This algorithm may result in the timer being set in the past, especially if timers wake up late. Timers set in the past fire immediately. Wild, I know 🙈

Pseudocode for SetLossDetectionTimer follows (where the "^" operator represents exponentiation):

   GetLossTimeAndSpace():
     time = loss_time[Initial]
     space = Initial
     for pn_space in [ Handshake, ApplicationData ]:
       if (time == 0 || loss_time[pn_space] < time):
         time = loss_time[pn_space];
         space = pn_space
     return time, space

   GetPtoTimeAndSpace():
     duration = (smoothed_rtt + max(4 * rttvar, kGranularity))
         * (2 ^ pto_count)
     // Anti-deadlock PTO starts from the current time
     if (no ack-eliciting packets in flight):
       assert(!PeerCompletedAddressValidation())
       if (has handshake keys):
         return (now() + duration), Handshake
       else:
         return (now() + duration), Initial
     pto_timeout = infinite
     pto_space = Initial
     for space in [ Initial, Handshake, ApplicationData ]:
       if (no ack-eliciting packets in flight in space):
           continue;
       if (space == ApplicationData):
         // Skip Application Data until handshake confirmed.
         if (handshake is not confirmed):
           return pto_timeout, pto_space
         // Include max_ack_delay and backoff for Application Data.
         duration += max_ack_delay * (2 ^ pto_count)

       t = time_of_last_ack_eliciting_packet[space] + duration
       if (t < pto_timeout):
         pto_timeout = t
         pto_space = space
     return pto_timeout, pto_space

   PeerCompletedAddressValidation():
     // Assume clients validate the server's address implicitly.
     if (endpoint is server):
       return true
     // Servers complete address validation when a
     // protected packet is received.
     return has received Handshake ACK ||
          handshake confirmed

   SetLossDetectionTimer():
     earliest_loss_time, _ = GetLossTimeAndSpace()
     if (earliest_loss_time != 0):
       // Time threshold loss detection.
       loss_detection_timer.update(earliest_loss_time)
       return

     if (server is at anti-amplification limit):
       // The server's timer is not set if nothing can be sent.
       loss_detection_timer.cancel()
       return

     if (no ack-eliciting packets in flight &&
         PeerCompletedAddressValidation()):
       // There is nothing to detect lost, so no timer is set.
       // However, the client needs to arm the timer if the
       // server might be blocked by the anti-amplification limit.
       loss_detection_timer.cancel()
       return

     timeout, _ = GetPtoTimeAndSpace()
     loss_detection_timer.update(timeout)

A.9. On Timeout ⏰💥

When the loss detection timer expires, the timer's mode decides what happens next. Each mode gets its own vibe 🌈

Pseudocode for OnLossDetectionTimeout follows:

   OnLossDetectionTimeout():
     earliest_loss_time, pn_space = GetLossTimeAndSpace()
     if (earliest_loss_time != 0):
       // Time threshold loss Detection
       lost_packets = DetectAndRemoveLostPackets(pn_space)
       assert(!lost_packets.empty())
       OnPacketsLost(lost_packets)
       SetLossDetectionTimer()
       return

     if (no ack-eliciting packets in flight):
       assert(!PeerCompletedAddressValidation())
       // Client sends an anti-deadlock packet: Initial is padded
       // to earn more anti-amplification credit,
       // a Handshake packet proves address ownership.
       if (has Handshake keys):
         SendOneAckElicitingHandshakePacket()
       else:
         SendOneAckElicitingPaddedInitialPacket()
     else:
       // PTO. Send new data if available, else retransmit old data.
       // If neither is available, send a single PING frame.
       _, pn_space = GetPtoTimeAndSpace()
       SendOneOrTwoAckElicitingPackets(pn_space)

     pto_count++
     SetLossDetectionTimer()

A.10. Detecting Lost Packets ✨

okay so DetectAndRemoveLostPackets gets called every single time an ACK comes in, or when the time threshold loss detection timer goes off 🫶. It works on the sent_packets for that packet number space and hands back a list of packets that have just been detected as lost. literally the lost-packet detective 🕵️‍♀️

Here's the pseudocode, bestie 💖

DetectAndRemoveLostPackets(pn_space):
  assert(largest_acked_packet[pn_space] != infinite)
  loss_time[pn_space] = 0
  lost_packets = []
  loss_delay = kTimeThreshold * max(latest_rtt, smoothed_rtt)

  // Minimum time of kGranularity before packets are deemed lost.
  loss_delay = max(loss_delay, kGranularity)

  // Packets sent before this time are deemed lost.
  lost_send_time = now() - loss_delay

  foreach unacked in sent_packets[pn_space]:
    if (unacked.packet_number > largest_acked_packet[pn_space]):
      continue

    // Mark packet as lost, or set time when it should be marked.
    // Note: The use of kPacketThreshold here assumes that there
    // were no sender-induced gaps in the packet number space.
    if (unacked.time_sent <= lost_send_time ||
        largest_acked_packet[pn_space] >=
          unacked.packet_number + kPacketThreshold):
      sent_packets[pn_space].remove(unacked.packet_number)
      lost_packets.insert(unacked)
    else:
      if (loss_time[pn_space] == 0):
        loss_time[pn_space] = unacked.time_sent + loss_delay
      else:
        loss_time[pn_space] = min(loss_time[pn_space],
                                  unacked.time_sent + loss_delay)
  return lost_packets

A.11. Upon Dropping Initial or Handshake Keys 🔑

when the Initial or Handshake keys get thrown out, the packets from that space go too, and the loss detection state gets updated to match 🧹 it's giving clean slate energy.

Here's the pseudocode for OnPacketNumberSpaceDiscarded, iykyk 💅

OnPacketNumberSpaceDiscarded(pn_space):
  assert(pn_space != ApplicationData)
  RemoveFromBytesInFlight(sent_packets[pn_space])
  sent_packets[pn_space].clear()
  // Reset the loss detection and PTO timer
  time_of_last_ack_eliciting_packet[pn_space] = 0
  loss_time[pn_space] = 0
  pto_count = 0
  SetLossDetectionTimer()

Appendix B. Congestion Control Pseudocode 🌈

okay this part is the congestion control glow-up ✨ we're about to go through an example implementation of the congestion controller from Section 7. the vibe is so organised 💅

The pseudocode segments in this section are licensed as Code Components; see the copyright notice.

B.1. Constants of Interest 🎀

these constants come from a mix of RFCs, papers, and common practice, so they're basically the greatest hits 🎶

kInitialWindow: the default limit on the initial bytes in flight, as described in Section 7.2.
kMinimumWindow: the minimum congestion window in bytes, as described in Section 7.2.
kLossReductionFactor: the scaling factor used to shrink the congestion window when a new loss event is detected. Section 7 recommends a value of 0.5.
kPersistentCongestionThreshold: the period of time it takes for persistent congestion to be established, specified as a PTO multiplier. Section 7.6 recommends a value of 3.

B.2. Variables of Interest 🦋

these are the variables the congestion control mechanisms need to do their thing 🧚

max_datagram_size: the sender's current maximum payload size. This does not include UDP or IP overhead. It's used for congestion window computations. An endpoint sets this based on its Path Maximum Transmission Unit (PMTU; see Section 14.2 of [QUIC-TRANSPORT]), with a minimum value of 1200 bytes.
ecn_ce_counters[kPacketNumberSpace]: the highest value the peer has reported for the ECN-CE counter in the packet number space, in an ACK frame. It's how we spot increases in the reported ECN-CE counter 👀
bytes_in_flight: the sum of the size in bytes of all sent packets that contain at least one ack-eliciting or PADDING frame and haven't been acknowledged or declared lost. The size doesn't include IP or UDP overhead, but it does include the QUIC header and Authenticated Encryption with Associated Data (AEAD) overhead. Packets that only contain ACK frames don't count toward bytes_in_flight, so congestion control can't get in the way of congestion feedback. the way that's protecting the feedback loop 🫶
congestion_window: the maximum number of bytes allowed to be in flight.
congestion_recovery_start_time: the time the current recovery period started, because of loss or ECN being detected. Once a packet sent after this time gets acknowledged, QUIC exits congestion recovery 🌸
ssthresh: the slow start threshold in bytes. When the congestion window is below ssthresh, the mode is slow start and the window grows by the number of bytes acknowledged.

The congestion control pseudocode also reads some of the variables from the loss recovery pseudocode, so keep those in your back pocket 🎀

B.3. Initialization 🌱

at the very start of a connection, you set up the congestion control variables like this:

congestion_window = kInitialWindow
bytes_in_flight = 0
congestion_recovery_start_time = 0
ssthresh = infinite
for pn_space in [ Initial, Handshake, ApplicationData ]:
  ecn_ce_counters[pn_space] = 0

B.4. On Packet Sent 📤

whenever a packet goes out and it has non-ACK frames in it, the packet bumps up bytes_in_flight.

OnPacketSentCC(sent_bytes):
  bytes_in_flight += sent_bytes

B.5. On Packet Acknowledgment 💌

this one gets invoked from loss detection's OnAckReceived, and it's fed the newly acked_packets from sent_packets.

heads up: in congestion avoidance, implementers who use an integer representation for congestion_window should be careful with division. they can use the alternative approach suggested in Section 2.1 of [RFC3465]. be careful, bestie 🙈

InCongestionRecovery(sent_time):
  return sent_time <= congestion_recovery_start_time

OnPacketsAcked(acked_packets):
  for acked_packet in acked_packets:
    OnPacketAcked(acked_packet)

OnPacketAcked(acked_packet):
  if (!acked_packet.in_flight):
    return;
  // Remove from bytes_in_flight.
  bytes_in_flight -= acked_packet.sent_bytes
  // Do not increase congestion_window if application
  // limited or flow control limited.
  if (IsAppOrFlowControlLimited())
    return
  // Do not increase congestion window in recovery period.
  if (InCongestionRecovery(acked_packet.time_sent)):
    return
  if (congestion_window < ssthresh):
    // Slow start.
    congestion_window += acked_packet.sent_bytes
  else:
    // Congestion avoidance.
    congestion_window +=
      max_datagram_size * acked_packet.sent_bytes
      / congestion_window

B.6. On New Congestion Event 🚨

this gets invoked from ProcessECN and OnPacketsLost when a new congestion event shows up. if we're not already in recovery, it starts a recovery period and immediately cuts the slow start threshold and congestion window. dramatic, but it's the good kind of dramatic 💀

OnCongestionEvent(sent_time):
  // No reaction if already in a recovery period.
  if (InCongestionRecovery(sent_time)):
    return

  // Enter recovery period.
  congestion_recovery_start_time = now()
  ssthresh = congestion_window * kLossReductionFactor
  congestion_window = max(ssthresh, kMinimumWindow)
  // A packet can be sent to speed up loss recovery.
  MaybeSendOnePacket()

B.7. Process ECN Information 🧠

this gets invoked when an ACK frame with an ECN section arrives from the peer.

ProcessECN(ack, pn_space):
  // If the ECN-CE counter reported by the peer has increased,
  // this could be a new congestion event.
  if (ack.ce_counter > ecn_ce_counters[pn_space]):
    ecn_ce_counters[pn_space] = ack.ce_counter
    sent_time = sent_packets[ack.largest_acked].time_sent
    OnCongestionEvent(sent_time)

B.8. On Packets Lost 📉

this gets invoked when DetectAndRemoveLostPackets decides packets are lost. it's a lot, so let's go step by step 🫶

OnPacketsLost(lost_packets):
  sent_time_of_last_loss = 0
  // Remove lost packets from bytes_in_flight.
  for lost_packet in lost_packets:
    if lost_packet.in_flight:
      bytes_in_flight -= lost_packet.sent_bytes
      sent_time_of_last_loss =
        max(sent_time_of_last_loss, lost_packet.time_sent)
  // Congestion event if in-flight packets were lost
  if (sent_time_of_last_loss != 0):
    OnCongestionEvent(sent_time_of_last_loss)

  // Reset the congestion window if the loss of these
  // packets indicates persistent congestion.
  // Only consider packets sent after getting an RTT sample.
  if (first_rtt_sample == 0):
    return
  pc_lost = []
  for lost in lost_packets:
    if lost.time_sent > first_rtt_sample:
      pc_lost.insert(lost)
  if (InPersistentCongestion(pc_lost)):
    congestion_window = kMinimumWindow
    congestion_recovery_start_time = 0

B.9. Removing Discarded Packets from Bytes in Flight 🗑️

when Initial or Handshake keys get discarded, the packets sent in that space stop counting toward bytes in flight. bye bye, packets 👋

Here's the pseudocode for RemoveFromBytesInFlight:

RemoveFromBytesInFlight(discarded_packets):
  // Remove any unacknowledged packets from flight.
  foreach packet in discarded_packets:
    if packet.in_flight
      bytes_in_flight -= size

Contributors 💜

the IETF QUIC Working Group got an enormous amount of support from so many people 🌈 these folks made substantive contributions to this document:

Alessandro Ghedini
Benjamin Saunders
Gorry Fairhurst
山本和彦 (Kazu Yamamoto)
奥一穂 (Kazuho Oku)
Lars Eggert
Magnus Westerlund
Marten Seemann
Martin Duke
Martin Thomson
Mirja Kühlewind
Nick Banks
Praveen Balasubramanian

Authors' Addresses 📬

Jana Iyengar (editor) Fastly

Email: jri.ietf@gmail.com

Ian Swett (editor) Google

Email: ianswett@google.com

📦 where this page comes from

ietf: rfc9002 2021-05
💖 girlified by claude-haiku-5-5 on 2026-10-18: a language model's rewrite of the original, under its terms