Internet-Draft | QUIC event definitions for qlog | August 2022 |
Marx, et al. | Expires 4 March 2023 | [Page] |
This document describes concrete qlog event definitions and their metadata for QUIC events. These events can then be embedded in the higher level schema defined in [QLOG-MAIN].¶
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.¶
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.¶
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."¶
This Internet-Draft will expire on 4 March 2023.¶
Copyright (c) 2022 IETF Trust and the persons identified as the document authors. All rights reserved.¶
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 Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License.¶
This document describes the values of the qlog name ("category" + "event") and "data" fields and their semantics for the QUIC protocol. This document is based on draft-34 of the QUIC I-Ds [QUIC-TRANSPORT], [QUIC-RECOVERY], and [QUIC-TLS]. HTTP/3 and QPACK events are defined in a separate document [QLOG-H3].¶
Feedback and discussion are welcome at https://github.com/quicwg/qlog. Readers are advised to refer to the "editor's draft" at that URL for an up-to-date version of this document.¶
Concrete examples of integrations of this schema in various programming languages can be found at https://github.com/quiclog/qlog/.¶
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in [RFC2119].¶
The event and data structure definitions in ths document are expressed in the Concise Data Definition Language [CDDL] and its extensions described in [QLOG-MAIN].¶
This document describes the values of the qlog "name" ("category" + "event") and "data" fields and their semantics for the QUIC protocol.¶
This document assumes the usage of the encompassing main qlog schema defined in
[QLOG-MAIN]. Each subsection below defines a separate category (for example
connectivity, transport, recovery) and each subsubsection is an event type (for
example packet_received
).¶
For each event type, its importance and data definition is laid out, often accompanied by possible values for the optional "trigger" field. For the definition and semantics of "importance" and "trigger", see the main schema document.¶
Most of the complex datastructures, enums and re-usable definitions are grouped together on the bottom of this document for clarity.¶
This document re-uses all the fields defined in the main qlog schema (e.g., name, category, type, data, group_id, protocol_type, the time-related fields, importance, RawInfo, etc.).¶
One entry in the "protocol_type" qlog array field MUST be "QUIC" if events from this document are included in a qlog trace.¶
When the qlog "group_id" field is used, it is recommended to use QUIC's Original Destination Connection ID (ODCID, the CID chosen by the client when first contacting the server), as this is the only value that does not change over the course of the connection and can be used to link more advanced QUIC packets (e.g., Retry, Version Negotiation) to a given connection. Similarly, the ODCID should be used as the qlog filename or file identifier, potentially suffixed by the vantagepoint type (For example, abcd1234_server.qlog would contain the server-side trace of the connection with ODCID abcd1234).¶
This document re-uses the definition of the RawInfo data class from [QLOG-MAIN].¶
QUIC packets always include an AEAD authentication tag ("trailer") at the end. As this tag is always the same size for a given connection (it depends on the used TLS cipher), this document does not define a separate "RawInfo:aead_tag_length" field here. Instead, this field is reflected in "transport:parameters_set" and can be logged only once.¶
As QUIC uses trailers in packets, packet header_lengths can be calculated as:¶
header_length = length - payload_length - aead_tag_length¶
For UDP datagrams, the calulation is simpler:¶
header_length = length - payload_length¶
In some cases, the length fields are also explicitly reflected inside of packet headers. For example, the QUIC STREAM frame has a "length" field indicating its payload size. Similarly, the QUIC Long Header has a "length" field which is equal to the payload length plus the packet number length. In these cases, those fields are intentionally preserved in the event definitions. Even though this can lead to duplicate data when the full RawInfo is logged, it allows a more direct mapping of the QUIC specifications to qlog, making it easier for users to interpret.¶
For several types of events, it is sometimes impossible to tie them to a specific conceptual QUIC connection (e.g., a packet_dropped event triggered because the packet has an unknown connection_id in the header). Since qlog events in a trace are typically associated with a single connection, it is unclear how to log these events.¶
Ideally, implementers SHOULD create a separate, individual "endpoint-level" trace file (or group_id value), not associated with a specific connection (for example a "server.qlog" or group_id = "client"), and log all events that do not belong to a single connection to this grouping trace. However, this is not always practical, depending on the implementation. Because the semantics of most of these events are well-defined in the protocols and because they are difficult to mis-interpret as belonging to a connection, implementers MAY choose to log events not belonging to a particular connection in any other trace, even those strongly associated with a single connection.¶
Note that this can make it difficult to match logs from different vantage points with each other. For example, from the client side, it is easy to log connections with version negotiation or retry in the same trace, while on the server they would most likely be logged in separate traces. Servers can take extra efforts (and keep additional state) to keep these events combined in a single trace however (for example by also matching connections on their four-tuple instead of just the connection ID).¶
Each subheading in this section is a qlog event category, while each sub-subheading is a qlog event type. Concretely, for the following two items, we have the category "connectivity" and event type "server_listening", resulting in a concatenated qlog "name" field value of "connectivity:server_listening".¶
Importance: Extra¶
Emitted when the server starts accepting connections.¶
Definition:¶
Note: some QUIC stacks do not handle sockets directly and are thus unable to log IP and/or port information.¶
Importance: Base¶
Used for both attempting (client-perspective) and accepting (server-perspective) new connections. Note that this event has overlap with connection_state_updated and this is a separate event mainly because of all the additional data that should be logged.¶
Definition:¶
Note: some QUIC stacks do not handle sockets directly and are thus unable to log IP and/or port information.¶
Importance: Base¶
Used for logging when a connection was closed, typically when an error or timeout occurred. Note that this event has overlap with connectivity:connection_state_updated, as well as the CONNECTION_CLOSE frame. However, in practice, when analyzing large deployments, it can be useful to have a single event representing a connection_closed event, which also includes an additional reason field to provide additional information. Additionally, it is useful to log closures due to timeouts, which are difficult to reflect using the other options.¶
In QUIC there are two main connection-closing error categories: connection and application errors. They have well-defined error codes and semantics. Next to these however, there can be internal errors that occur that may or may not get mapped to the official error codes in implementation-specific ways. As such, multiple error codes can be set on the same event to reflect this.¶
Definition:¶
Importance: Base¶
This event is emitted when either party updates their current Connection ID. As this typically happens only sparingly over the course of a connection, this event allows loggers to be more efficient than logging the observed CID with each packet in the .header field of the "packet_sent" or "packet_received" events.¶
This is viewed from the perspective of the one applying the new id. As such, if we receive a new connection id from our peer, we will see the dst_ fields are set. If we update our own connection id (e.g., NEW_CONNECTION_ID frame), we log the src_ fields.¶
Definition:¶
Importance: Base¶
To be emitted when the spin bit changes value. It SHOULD NOT be emitted if the spin bit is set without changing its value.¶
Definition:¶
Importance: Base¶
This event is used to track progress through QUIC's complex handshake and connection close procedures. It is intended to provide exhaustive options to log each state individually, but also provides a more basic, simpler set for implementations less interested in tracking each smaller state transition. As such, users should not expect to see -all- these states reflected in all qlogs and implementers should focus on support for the SimpleConnectionState set.¶
Definition:¶
These states correspond to the following transitions for both client and server:¶
Client:¶
send initial¶
get initial¶
get first Handshake packet¶
get Handshake packet containing ServerFinished¶
send ClientFinished¶
get HANDSHAKE_DONE¶
Server:¶
get initial¶
send handshake EE, CERT, CV, ...¶
send ServerFinished¶
get first handshake packet / something using a server-issued CID of min length¶
get handshake packet containing ClientFinished¶
send HANDSHAKE_DONE¶
connection_state_changed with a new state of "attempted" is the same conceptual event as the connection_started event above from the client's perspective. Similarly, a state of "closing" or "draining" corresponds to the connection_closed event.¶
Importance: Base¶
Note: secret_updated would be more correct, but in the draft it's called KEY_UPDATE, so stick with that for consistency¶
Definition:¶
Importance: Base¶
Definition:¶
Importance: Core¶
QUIC endpoints each have their own list of of QUIC versions they support. The client uses the most likely version in their first initial. If the server does support that version, it replies with a version_negotiation packet, containing supported versions. From this, the client selects a version. This event aggregates all this information in a single event type. It also allows logging of supported versions at an endpoint without actual version negotiation needing to happen.¶
Definition:¶
Intended use:¶
Importance: Core¶
QUIC implementations each have their own list of application level protocols and versions thereof they support. The client includes a list of their supported options in its first initial as part of the TLS Application Layer Protocol Negotiation (alpn) extension. If there are common option(s), the server chooses the most optimal one and communicates this back to the client. If not, the connection is closed.¶
Definition:¶
Intended use:¶
Importance: Core¶
This event groups settings from several different sources (transport parameters, TLS ciphers, etc.) into a single event. This is done to minimize the amount of events and to decouple conceptual setting impacts from their underlying mechanism for easier high-level reasoning.¶
All these settings are typically set once and never change. However, they are typically set at different times during the connection, so there will typically be several instances of this event with different fields set.¶
Note that some settings have two variations (one set locally, one requested by the remote peer). This is reflected in the "owner" field. As such, this field MUST be correct for all settings included a single event instance. If you need to log settings from two sides, you MUST emit two separate event instances.¶
In the case of connection resumption and 0-RTT, some of the server's parameters
are stored up-front at the client and used for the initial connection startup.
They are later updated with the server's reply. In these cases, utilize the
separate parameters_restored
event to indicate the initial values, and this
event to indicate the updated values, as normal.¶
Definition:¶
Additionally, this event can contain any number of unspecified fields. This is to reflect setting of for example unknown (greased) transport parameters or employed (proprietary) extensions.¶
Importance: Base¶
When using QUIC 0-RTT, clients are expected to remember and restore the server's transport parameters from the previous connection. This event is used to indicate which parameters were restored and to which values when utilizing 0-RTT. Note that not all transport parameters should be restored (many are even prohibited from being re-utilized). The ones listed here are the ones expected to be useful for correct 0-RTT usage.¶
Definition:¶
Note that, like parameters_set above, this event can contain any number of unspecified fields to allow for additional/custom parameters.¶
Importance: Core¶
Definition:¶
Note: We do not explicitly log the encryption_level or packet_number_space: the header.packet_type specifies this by inference (assuming correct implementation)¶
Note: for more details on "datagram_id", see Section 3.3.10. It is only needed when keeping track of packet coalescing.¶
Importance: Core¶
Definition:¶
Note: We do not explicitly log the encryption_level or packet_number_space: the header.packet_type specifies this by inference (assuming correct implementation)¶
Note: for more details on "datagram_id", see Section 3.3.10. It is only needed when keeping track of packet coalescing.¶
Importance: Base¶
This event indicates a QUIC-level packet was dropped after partial or no parsing.¶
Definition:¶
Note: sometimes packets are dropped before they can be associated with a particular connection (e.g., in case of "unsupported_version"). This situation is discussed more in Section 2.1.2.¶
Note: for more details on "datagram_id", see Section 3.3.10. It is only needed when keeping track of packet coalescing.¶
Importance: Base¶
This event is emitted when a packet is buffered because it cannot be processed yet. Typically, this is because the packet cannot be parsed yet, and thus we only log the full packet contents when it was parsed in a packet_received event.¶
Definition:¶
Note: for more details on "datagram_id", see Section 3.3.10. It is only needed when keeping track of packet coalescing.¶
Importance: Extra¶
This event is emitted when a (group of) sent packet(s) is acknowledged by the remote peer for the first time. This information could also be deduced from the contents of received ACK frames. However, ACK frames require additional processing logic to determine when a given packet is acknowledged for the first time, as QUIC uses ACK ranges which can include repeated ACKs. Additionally, this event can be used by implementations that do not log frame contents.¶
Definition:¶
Note: if packet_number_space is omitted, it assumes the default value of PacketNumberSpace.application_data, as this is by far the most prevalent packet number space a typical QUIC connection will use.¶
Importance: Extra¶
When we pass one or more UDP-level datagrams to the socket. This is useful for determining how QUIC packet buffers are drained to the OS.¶
Definition:¶
Note: QUIC itself does not have a concept of a "datagram_id". This field is a purely qlog-specific construct to allow tracking how multiple QUIC packets are coalesced inside of a single UDP datagram, which is an important optimization during the QUIC handshake. For this, implementations assign a (per-endpoint) unique ID to each datagram and keep track of which packets were coalesced into the same datagram. As packet coalescing typically only happens during the handshake (as it requires at least one long header packet), this can be done without much overhead.¶
Importance: Extra¶
When we receive one or more UDP-level datagrams from the socket. This is useful for determining how datagrams are passed to the user space stack from the OS.¶
Definition:¶
Note: for more details on "datagram_ids", see Section 3.3.10.¶
Importance: Extra¶
When we drop a UDP-level datagram. This is typically if it does not contain a valid QUIC packet (in that case, use packet_dropped instead).¶
Definition:¶
Importance: Base¶
This event is emitted whenever the internal state of a QUIC stream is updated, as described in QUIC transport draft-23 section 3. Most of this can be inferred from several types of frames going over the wire, but it's much easier to have explicit signals for these state changes.¶
Definition:¶
Note: QUIC implementations SHOULD mainly log the simplified bidirectional (HTTP/2-alike) stream states (e.g., idle, open, closed) instead of the more finegrained stream states (e.g., data_sent, reset_received). These latter ones are mainly for more in-depth debugging. Tools SHOULD be able to deal with both types equally.¶
Importance: Extra¶
This event's main goal is to prevent a large proliferation of specific purpose events (e.g., packets_acknowledged, flow_control_updated, stream_data_received). We want to give implementations the opportunity to (selectively) log this type of signal without having to log packet-level details (e.g., in packet_received). Since for almost all cases, the effects of applying a frame to the internal state of an implementation can be inferred from that frame's contents, we aggregate these events in this single "frames_processed" event.¶
Note: This event can be used to signal internal state change not resulting directly from the actual "parsing" of a frame (e.g., the frame could have been parsed, data put into a buffer, then later processed, then logged with this event).¶
Note: Implementations logging "packet_received" and which include all of the packet's constituent frames therein, are not expected to emit this "frames_processed" event. Rather, implementations not wishing to log full packets or that wish to explicitly convey extra information about when frames are processed (if not directly tied to their reception) can use this event.¶
Note: for some events, this approach will lose some information (e.g., for which encryption level are packets being acknowledged?). If this information is important, please use the packet_received event instead.¶
Note: in some implementations, it can be difficult to log frames directly, even when using packet_sent and packet_received events. For these cases, this event also contains the direct packet_number field, which can be used to more explicitly link this event to the packet_sent/received events.¶
Definition:¶
Importance: Base¶
Used to indicate when data moves between the different layers (for example passing from the application protocol (e.g., HTTP) to QUIC stream buffers and vice versa) or between the application protocol (e.g., HTTP) and the actual user application on top (for example a browser engine). This helps make clear the flow of data, how long data remains in various buffers and the overheads introduced by individual layers.¶
For example, this helps make clear whether received data on a QUIC stream is moved to the application protocol immediately (for example per received packet) or in larger batches (for example, all QUIC packets are processed first and afterwards the application layer reads from the streams with newly available data). This in turn can help identify bottlenecks or scheduling problems.¶
Definition:¶
Note: we do not for example use a "direction" field (with values "up" and "down") to specify the data flow. This is because in some optimized implementations, data might skip some individual layers. Additionally, using explicit "from" and "to" fields is more flexible and allows the definition of other conceptual "layers" (for example to indicate data from QUIC CRYPTO frames being passed to a TLS library ("security") or from HTTP/3 to QPACK ("qpack")).¶
Note: this event type is part of the "transport" category, but really spans all the different layers. This means we have a few leaky abstractions here (for example, the stream_id or stream offset might not be available at some logging points, or the raw data might not be in a byte-array form). In these situations, implementers can decide to define new, in-context fields to aid in manual debugging.¶
Note: most of the events in this category are kept generic to support different recovery approaches and various congestion control algorithms. Tool creators SHOULD make an effort to support and visualize even unknown data in these events (e.g., plot unknown congestion states by name on a timeline visualization).¶
Importance: Base¶
This event groups initial parameters from both loss detection and congestion control into a single event. All these settings are typically set once and never change. Implementation that do, for some reason, change these parameters during execution, MAY emit the parameters_set event twice.¶
Definition:¶
Additionally, this event can contain any number of unspecified fields to support different recovery approaches.¶
Importance: Core¶
This event is emitted when one or more of the observable recovery metrics changes value. This event SHOULD group all possible metric updates that happen at or around the same time in a single event (e.g., if min_rtt and smoothed_rtt change at the same time, they should be bundled in a single metrics_updated entry, rather than split out into two). Consequently, a metrics_updated event is only guaranteed to contain at least one of the listed metrics.¶
Definition:¶
Note: to make logging easier, implementations MAY log values even if they are the same as previously reported values (e.g., two subsequent RecoveryMetricsUpdated entries can both report the exact same value for min_rtt). However, applications SHOULD try to log only actual updates to values.¶
Additionally, this event can contain any number of unspecified fields to support different recovery approaches.¶
Importance: Base¶
This event signifies when the congestion controller enters a significant new state and changes its behaviour. This event's definition is kept generic to support different Congestion Control algorithms. For example, for the algorithm defined in the Recovery draft ("enhanced" New Reno), the following states are defined:¶
Definition:¶
The "trigger" field SHOULD be logged if there are multiple ways in which a state change can occur but MAY be omitted if a given state can only be due to a single event occuring (e.g., slow start is exited only when ssthresh is exceeded).¶
Importance: Extra¶
This event is emitted when a recovery loss timer changes state. The three main event types are:¶
Note: to indicate an active timer's timeout update, a new "set" event is used.¶
Definition:¶
TODO: how about CC algo's that use multiple timers? How generic do these events need to be? Just support QUIC-style recovery from the spec or broader?¶
TODO: read up on the loss detection logic in draft-27 onward and see if this suffices¶
Importance: Core¶
This event is emitted when a packet is deemed lost by loss detection.¶
Definition:¶
For this event, the "trigger" field SHOULD be set (for example to one of the values below), as this helps tremendously in debugging.¶
Importance: Extra¶
This event indicates which data was marked for retransmit upon detecing a packet loss (see packet_lost). Similar to our reasoning for the "frames_processed" event, in order to keep the amount of different events low, we group this signal for all types of retransmittable data in a single event based on existing QUIC frame definitions.¶
Implementations retransmitting full packets or frames directly can just log the consituent frames of the lost packet here (or do away with this event and use the contents of the packet_lost event instead). Conversely, implementations that have more complex logic (e.g., marking ranges in a stream's data buffer as in-flight), or that do not track sent frames in full (e.g., only stream offset + length), can translate their internal behaviour into the appropriate frame instance here even if that frame was never or will never be put on the wire.¶
Note: much of this data can be inferred if implementations log packet_sent events (e.g., looking at overlapping stream data offsets and length, one can determine when data was retransmitted).¶
Definition:¶
We extend the $ProtocolEventBody
extension point defined in
[QLOG-MAIN] with the QUIC protocol events defined in this document.¶
QuicEvents = ConnectivityServerListening / ConnectivityConnectionStarted / ConnectivityConnectionClosed / ConnectivityConnectionIDUpdated / ConnectivitySpinBitUpdated / ConnectivityConnectionStateUpdated / SecurityKeyUpdated / SecurityKeyRetired / TransportVersionInformation / TransportALPNInformation / TransportParametersSet / TransportParametersRestored / TransportPacketSent / TransportPacketReceived / TransportPacketDropped / TransportPacketBuffered / TransportPacketsAcked / TransportDatagramsSent / TransportDatagramsReceived / TransportDatagramDropped / TransportStreamStateUpdated / TransportFramesProcessed / TransportDataMoved / RecoveryParametersSet / RecoveryMetricsUpdated / RecoveryCongestionStateUpdated / RecoveryLossTimerUpdated / RecoveryPacketLost $ProtocolEventBody /= QuicEvents¶
The token carried in an Initial packet can either be a retry token from a Retry packet, or one originally provided by the server in a NEW_TOKEN frame used when resuming a connection (e.g., for address validation purposes). Retry and resumption tokens typically contain encoded metadata to check the token's validity when it is used, but this metadata and its format is implementation specific. For that, this field includes a general-purpose "details" field.¶
The stateless reset token is carried in stateless reset packets, in transport parameters and in NEW_CONNECTION_ID frames.¶
In QUIC, PADDING frames are simply identified as a single byte of value 0. As such, each padding byte could be theoretically interpreted and logged as an individual PaddingFrame.¶
However, as this leads to heavy logging overhead, implementations SHOULD instead emit just a single PaddingFrame and set the payload_length property to the amount of PADDING bytes/frames included in the packet.¶
Note: the packet ranges in AckFrame.acked_ranges do not necessarily have to be ordered (e.g., [[5,9],[1,4]] is a valid value).¶
Note: the two numbers in the packet range can be the same (e.g., [120,120] means that packet with number 120 was ACKed). However, in that case, implementers SHOULD log [120] instead and tools MUST be able to deal with both notations.¶
raw_error_code is the actual, numerical code. This is useful because some error types are spread out over a range of codes (e.g., QUIC's crypto_error).¶
By definition, an application error is defined by the application-level protocol running on top of QUIC (e.g., HTTP/3).¶
As such, we cannot define it here directly. Though we provide an extension point through the use of the CDDL "socket" mechanism.¶
Application-level qlog definitions that wish to define new ApplicationError strings MUST do so by extending the $ApplicationError socket as such:¶
$ApplicationError /= "new_error_name" / "another_new_error_name"¶
These errors are defined in the TLS document as "A TLS alert is turned into a QUIC connection error by converting the one-byte alert description into a QUIC error code. The alert description is added to 0x100 to produce a QUIC error code from the range reserved for CRYPTO_ERROR."¶
This approach maps badly to a pre-defined enum. As such, we define the crypto_error string as having a dynamic component here, which should include the hex-encoded and zero-padded value of the TLS alert description.¶
data_moved
event¶
Major changes:¶
Smaller changes:¶
Much of the initial work by Robin Marx was done at Hasselt University.¶
Thanks to Marten Seemann, Jana Iyengar, Brian Trammell, Dmitri Tikhonov, Stephen Petrides, Jari Arkko, Marcus Ihlar, Victor Vasiliev, Mirja Kuehlewind, Jeremy Laine, Kazu Yamamoto, Christian Huitema, and Lucas Pardue for their feedback and suggestions.¶