Internet-Draft DTLS 1.3 November 2020
Rescorla, et al. Expires 6 May 2021 [Page]
6347 (if approved)
Intended Status:
Standards Track
E. Rescorla
RTFM, Inc.
H. Tschofenig
Arm Limited
N. Modadugu
Google, Inc.

The Datagram Transport Layer Security (DTLS) Protocol Version 1.3


This document specifies Version 1.3 of the Datagram Transport Layer Security (DTLS) protocol. DTLS 1.3 allows client/server applications to communicate over the Internet in a way that is designed to prevent eavesdropping, tampering, and message forgery.

The DTLS 1.3 protocol is intentionally based on the Transport Layer Security (TLS) 1.3 protocol and provides equivalent security guarantees with the exception of order protection/non-replayability. Datagram semantics of the underlying transport are preserved by the DTLS protocol.

Status of This Memo

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

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 6 May 2021.

Table of Contents

1. Introduction


The source for this draft is maintained in GitHub. Suggested changes should be submitted as pull requests at Instructions are on that page as well. Editorial changes can be managed in GitHub, but any substantive change should be discussed on the TLS mailing list.

The primary goal of the TLS protocol is to establish an authenticated, confidentiality and integrity protected channel between two communicating peers. The TLS protocol is composed of two layers: the TLS Record Protocol and the TLS Handshake Protocol. However, TLS must run over a reliable transport channel - typically TCP [RFC0793].

There are applications that use UDP [RFC0768] as a transport and to offer communication security protection for those applications the Datagram Transport Layer Security (DTLS) protocol has been developed. DTLS is deliberately designed to be as similar to TLS as possible, both to minimize new security invention and to maximize the amount of code and infrastructure reuse.

DTLS 1.0 [RFC4347] was originally defined as a delta from TLS 1.1 [RFC4346] and DTLS 1.2 [RFC6347] was defined as a series of deltas to TLS 1.2 [RFC5246]. There is no DTLS 1.1; that version number was skipped in order to harmonize version numbers with TLS. This specification describes the most current version of the DTLS protocol based on TLS 1.3 [TLS13].

Implementations that speak both DTLS 1.2 and DTLS 1.3 can interoperate with those that speak only DTLS 1.2 (using DTLS 1.2 of course), just as TLS 1.3 implementations can interoperate with TLS 1.2 (see Appendix D of [TLS13] for details). While backwards compatibility with DTLS 1.0 is possible the use of DTLS 1.0 is not recommended as explained in Section 3.1.2 of RFC 7525 [RFC7525].

2. Conventions and Terminology

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.

The following terms are used:

The reader is assumed to be familiar with the TLS 1.3 specification since this document is defined as a delta from TLS 1.3. As in TLS 1.3 the HelloRetryRequest has the same format as a ServerHello message but for convenience we use the term HelloRetryRequest throughout this document as if it were a distinct message.

The reader is also as to be familiar with [I-D.ietf-tls-dtls-connection-id] as this document applies the CID functionality to DTLS 1.3.

Figures in this document illustrate various combinations of the DTLS protocol exchanges and the symbols have the following meaning:

3. DTLS Design Rationale and Overview

The basic design philosophy of DTLS is to construct "TLS over datagram transport". Datagram transport does not require nor provide reliable or in-order delivery of data. The DTLS protocol preserves this property for application data. Applications such as media streaming, Internet telephony, and online gaming use datagram transport for communication due to the delay-sensitive nature of transported data. The behavior of such applications is unchanged when the DTLS protocol is used to secure communication, since the DTLS protocol does not compensate for lost or reordered data traffic.

TLS cannot be used directly in datagram environments for the following five reasons:

  1. TLS relies on an implicit sequence number on records. If a record is not received, then the recipient will use the wrong sequence number when attempting to remove record protection from subsequent records. DTLS solves this problem by adding sequence numbers.
  2. The TLS handshake is a lock-step cryptographic handshake. Messages must be transmitted and received in a defined order; any other order is an error. DTLS handshake messages are also assigned sequence numbers to enable reassembly in the correct order in case datagrams are lost or reordered.
  3. During the handshake, messages are implicitly acknowledged by other handshake messages. Some handshake messages, such as the NewSessionTicket message, do not result in any direct response that would allow the sender to detect loss. DTLS adds an acknowledgment message to enable better loss recovery.
  4. Handshake messages are potentially larger than can be contained in a single datagram. DTLS adds fields to handshake messages to support fragmentation and reassembly.
  5. Datagram transport protocols, like UDP, are susceptible to abusive behavior effecting denial of service attacks against nonparticipants. DTLS adds a return-routability check that uses the TLS HelloRetryRequest message (see Section 5.1 for details).

3.1. Packet Loss

DTLS uses a simple retransmission timer to handle packet loss. Figure 1 demonstrates the basic concept, using the first phase of the DTLS handshake:

         Client                                   Server
         ------                                   ------
         ClientHello           ------>

                                 X<-- HelloRetryRequest

         [Timer Expires]

         ClientHello           ------>
Figure 1: DTLS retransmission example

Once the client has transmitted the ClientHello message, it expects to see a HelloRetryRequest or a ServerHello from the server. However, if the server's message is lost, the client knows that either the ClientHello or the response from the server has been lost and retransmits. When the server receives the retransmission, it knows to retransmit.

The server also maintains a retransmission timer and retransmits when that timer expires.

Note that timeout and retransmission do not apply to the HelloRetryRequest since this would require creating state on the server. The HelloRetryRequest is designed to be small enough that it will not itself be fragmented, thus avoiding concerns about interleaving multiple HelloRetryRequests.

3.2. Reordering

In DTLS, each handshake message is assigned a specific sequence number. When a peer receives a handshake message, it can quickly determine whether that message is the next message it expects. If it is, then it processes it. If not, it queues it for future handling once all previous messages have been received.

3.3. Message Size

TLS and DTLS handshake messages can be quite large (in theory up to 2^24-1 bytes, in practice many kilobytes). By contrast, UDP datagrams are often limited to less than 1500 bytes if IP fragmentation is not desired. In order to compensate for this limitation, each DTLS handshake message may be fragmented over several DTLS records, each of which is intended to fit in a single UDP datagram. Each DTLS handshake message contains both a fragment offset and a fragment length. Thus, a recipient in possession of all bytes of a handshake message can reassemble the original unfragmented message.

3.4. Replay Detection

DTLS optionally supports record replay detection. The technique used is the same as in IPsec AH/ESP, by maintaining a bitmap window of received records. Records that are too old to fit in the window and records that have previously been received are silently discarded. The replay detection feature is optional, since packet duplication is not always malicious, but can also occur due to routing errors. Applications may conceivably detect duplicate packets and accordingly modify their data transmission strategy.

4. The DTLS Record Layer

The DTLS 1.3 record layer is different from the TLS 1.3 record layer and also different from the DTLS 1.2 record layer.

  1. The DTLSCiphertext structure omits the superfluous version number and type fields.
  2. DTLS adds an epoch and sequence number to the TLS record header. This sequence number allows the recipient to correctly verify the DTLS MAC. However, the number of bits used for the epoch and sequence number fields in the DTLSCiphertext structure have been reduced from those in previous versions.
  3. The DTLSCiphertext structure has a variable length header.

DTLSPlaintext records are used to send unprotected records and DTLSCiphertext records are used to send protected records.

The DTLS record formats are shown below. Unless explicitly stated the meaning of the fields is unchanged from previous TLS / DTLS versions.

    struct {
        ContentType type;
        ProtocolVersion legacy_record_version;
        uint16 epoch = 0
        uint48 sequence_number;
        uint16 length;
        opaque fragment[DTLSPlaintext.length];
    } DTLSPlaintext;

    struct {
         opaque content[DTLSPlaintext.length];
         ContentType type;
         uint8 zeros[length_of_padding];
    } DTLSInnerPlaintext;

    struct {
        opaque unified_hdr[variable];
        opaque encrypted_record[length];
    } DTLSCiphertext;
Figure 2: DTLS 1.3 Record Format
The unified_hdr is a field of variable length, as shown in Figure 3.
Identical to the encrypted_record field in a TLS 1.3 record.

The DTLSCiphertext header is tightly bit-packed, as shown below:

    0 1 2 3 4 5 6 7
    |0|0|1|C|S|L|E E|
    | Connection ID |   Legend:
    | (if any,      |
    /  length as    /   C   - Connection ID (CID) present
    |  negotiated)  |   S   - Sequence number length
    +-+-+-+-+-+-+-+-+   L   - Length present
    |  8 or 16 bit  |   E   - Epoch
    |Sequence Number|
    | 16 bit Length |
    | (if present)  |
Figure 3: DTLS 1.3 CipherText Header
Fixed Bits:
The three high bits of the first byte of the DTLSCiphertext header are set to 001.
The C bit (0x10) is set if the Connection ID is present.
The S bit (0x08) indicates the size of the sequence number. 0 means an 8-bit sequence number, 1 means 16-bit.
The L bit (0x04) is set if the length is present.
The two low bits (0x03) include the low order two bits of the epoch.
Connection ID:
Variable length CID. The CID functionality is described in [I-D.ietf-tls-dtls-connection-id]. An example can be found in Section 9.1.
Sequence Number:
The low order 8 or 16 bits of the record sequence number. This value is 16 bits if the S bit is set to 1, and 8 bits if the S bit is 0.
Identical to the length field in a TLS 1.3 record.

As with previous versions of DTLS, multiple DTLSPlaintext and DTLSCiphertext records can be included in the same underlying transport datagram.

Figure 4 illustrates different record layer header types.

 0 1 2 3 4 5 6 7       0 1 2 3 4 5 6 7       0 1 2 3 4 5 6 7
+-+-+-+-+-+-+-+-+     +-+-+-+-+-+-+-+-+     +-+-+-+-+-+-+-+-+
| Content Type  |     |0|0|1|1|1|1|E E|     |0|0|1|0|0|0|E E|
+-+-+-+-+-+-+-+-+     +-+-+-+-+-+-+-+-+     +-+-+-+-+-+-+-+-+
|   16 bit      |     |               |     |8-bit Seq. No. |
|   Version     |     / Connection ID /     +-+-+-+-+-+-+-+-+
+-+-+-+-+-+-+-+-+     |               |     |               |
|   16 bit      |     +-+-+-+-+-+-+-+-+     |   Encrypted   |
|    Epoch      |     |    16 bit     |     /   Record      /
+-+-+-+-+-+-+-+-+     |Sequence Number|     |               |
|               |     +-+-+-+-+-+-+-+-+     +-+-+-+-+-+-+-+-+
|               |     |   16 bit      |
|   48 bit      |     |   Length      |       DTLSCiphertext
|Sequence Number|     +-+-+-+-+-+-+-+-+         Structure
|               |     |               |         (minimal)
|               |     |  Encrypted    |
+-+-+-+-+-+-+-+-+     /  Record       /
|    16 bit     |     |               |
|    Length     |     +-+-+-+-+-+-+-+-+
|               |      DTLSCiphertext
|               |        Structure
/   Fragment    /          (full)
|               |

Figure 4: Header Examples

The length field MAY be omitted by clearing the L bit, which means that the record consumes the entire rest of the datagram in the lower level transport. In this case it is not possible to have multiple DTLSCiphertext format records without length fields in the same datagram. Omitting the length field MUST only be used for the last record in a datagram.

If a connection ID is negotiated, then it MUST be contained in all datagrams. Sending implementations MUST NOT mix records from multiple DTLS associations in the same datagram. If the second or later record has a connection ID which does not correspond to the same association used for previous records, the rest of the datagram MUST be discarded.

When expanded, the epoch and sequence number can be combined into an unpacked RecordNumber structure, as shown below:

    struct {
        uint16 epoch;
        uint48 sequence_number;
    } RecordNumber;

This 64-bit value is used in the ACK message as well as in the "record_sequence_number" input to the AEAD function.

The entire header value shown in Figure 4 (but prior to record number encryption) is used as as the additional data value for the AEAD function. For instance, if the minimal variant is used, the AAD is 2 octets long. Note that this design is different from the additional data calculation for DTLS 1.2 and for DTLS 1.2 with Connection ID.

4.1. Determining the Header Format

Implementations can distinguish the two header formats by examining the first byte:

  • If the first byte is alert(21), handshake(22), or ack(proposed, 26), the record MUST be interpreted as a DTLSPlaintext record.
  • If the first byte is any other value, then receivers MUST check to see if the leading bits of the first byte are 001. If so, the implementation MUST process the record as DTLSCiphertext; the true content type will be inside the protected portion.
  • Otherwise, the record MUST be rejected as if it had failed deprotection, as described in Section 4.5.2.

4.2. Sequence Number and Epoch

DTLS uses an explicit or partly explicit sequence number, rather than an implicit one, carried in the sequence_number field of the record. Sequence numbers are maintained separately for each epoch, with each sequence_number initially being 0 for each epoch.

The epoch number is initially zero and is incremented each time keying material changes and a sender aims to rekey. More details are provided in Section 6.1.

4.2.1. Processing Guidelines

Because DTLS records could be reordered, a record from epoch M may be received after epoch N (where N > M) has begun. In general, implementations SHOULD discard records from earlier epochs, but if packet loss causes noticeable problems implementations MAY choose to retain keying material from previous epochs for up to the default MSL specified for TCP [RFC0793] to allow for packet reordering. (Note that the intention here is that implementers use the current guidance from the IETF for MSL, as specified in [RFC0793] or successors not that they attempt to interrogate the MSL that the system TCP stack is using.)

Conversely, it is possible for records that are protected with the new epoch to be received prior to the completion of a handshake. For instance, the server may send its Finished message and then start transmitting data. Implementations MAY either buffer or discard such records, though when DTLS is used over reliable transports (e.g., SCTP [RFC4960]), they SHOULD be buffered and processed once the handshake completes. Note that TLS's restrictions on when records may be sent still apply, and the receiver treats the records as if they were sent in the right order.

Implementations MUST send retransmissions of lost messages using the same epoch and keying material as the original transmission.

Implementations MUST either abandon an association or re-key prior to allowing the sequence number to wrap.

Implementations MUST NOT allow the epoch to wrap, but instead MUST establish a new association, terminating the old association.

4.2.2. Reconstructing the Sequence Number and Epoch

When receiving protected DTLS records message, the recipient does not have a full epoch or sequence number value and so there is some opportunity for ambiguity. Because the full epoch and sequence number are used to compute the per-record nonce, failure to reconstruct these values leads to failure to deprotect the record, and so implementations MAY use a mechanism of their choice to determine the full values. This section provides an algorithm which is comparatively simple and which implementations are RECOMMENDED to follow.

If the epoch bits match those of the current epoch, then implementations SHOULD reconstruct the sequence number by computing the full sequence number which is numerically closest to one plus the sequence number of the highest successfully deprotected record.

During the handshake phase, the epoch bits unambiguously indicate the correct key to use. After the handshake is complete, if the epoch bits do not match those from the current epoch implementations SHOULD use the most recent past epoch which has matching bits, and then reconstruct the sequence number as described above.

4.2.3. Sequence Number Encryption

In DTLS 1.3, when records are encrypted, record sequence numbers are also encrypted. The basic pattern is that the underlying encryption algorithm used with the AEAD algorithm is used to generate a mask which is then XORed with the sequence number.

When the AEAD is based on AES, then the Mask is generated by computing AES-ECB on the first 16 bytes of the ciphertext:

  Mask = AES-ECB(sn_key, Ciphertext[0..15])

When the AEAD is based on ChaCha20, then the mask is generated by treating the first 4 bytes of the ciphertext as the block counter and the next 12 bytes as the nonce, passing them to the ChaCha20 block function (Section 2.3 of [CHACHA]):

  Mask = ChaCha20(sn_key, Ciphertext[0..3], Ciphertext[4..15])

The sn_key is computed as follows:

   [sender]_sn_key  = HKDF-Expand-Label(Secret, "sn" , "", key_length)

[sender] denotes the sending side. The Secret value to be used is described in Section 7.3 of [TLS13].

The encrypted sequence number is computed by XORing the leading bytes of the Mask with the sequence number. Decryption is accomplished by the same process.

This procedure requires the ciphertext length be at least 16 bytes. Receivers MUST reject shorter records as if they had failed deprotection, as described in Section 4.5.2. Senders MUST pad short plaintexts out (using the conventional record padding mechanism) in order to make a suitable-length ciphertext. Note most of the DTLS AEAD algorithms have a 16-byte authentication tag and need no padding. However, some algorithms such as TLS_AES_128_CCM_8_SHA256 have a shorter authentication tag and may require padding for short inputs.

Note that sequence number encryption is only applied to the DTLSCiphertext structure and not to the DTLSPlaintext structure, which also contains a sequence number.

4.3. Transport Layer Mapping

DTLS messages MAY be fragmented into multiple DTLS records. Each DTLS record MUST fit within a single datagram. In order to avoid IP fragmentation, clients of the DTLS record layer SHOULD attempt to size records so that they fit within any PMTU estimates obtained from the record layer.

Multiple DTLS records MAY be placed in a single datagram. Records are encoded consecutively. The length field from DTLS records containing that field can be used to determine the boundaries between records. The final record in a datagram can omit the length field. The first byte of the datagram payload MUST be the beginning of a record. Records MUST NOT span datagrams.

DTLS records without CIDs do not contain any association identifiers and applications must arrange to multiplex between associations. With UDP, the host/port number is used to look up the appropriate security association for incoming records.

Some transports, such as DCCP [RFC4340], provide their own sequence numbers. When carried over those transports, both the DTLS and the transport sequence numbers will be present. Although this introduces a small amount of inefficiency, the transport layer and DTLS sequence numbers serve different purposes; therefore, for conceptual simplicity, it is superior to use both sequence numbers.

Some transports provide congestion control for traffic carried over them. If the congestion window is sufficiently narrow, DTLS handshake retransmissions may be held rather than transmitted immediately, potentially leading to timeouts and spurious retransmission. When DTLS is used over such transports, care should be taken not to overrun the likely congestion window. [RFC5238] defines a mapping of DTLS to DCCP that takes these issues into account.

4.4. PMTU Issues

In general, DTLS's philosophy is to leave PMTU discovery to the application. However, DTLS cannot completely ignore PMTU for three reasons:

  • The DTLS record framing expands the datagram size, thus lowering the effective PMTU from the application's perspective.
  • In some implementations, the application may not directly talk to the network, in which case the DTLS stack may absorb ICMP [RFC1191] "Datagram Too Big" indications or ICMPv6 [RFC4443] "Packet Too Big" indications.
  • The DTLS handshake messages can exceed the PMTU.

In order to deal with the first two issues, the DTLS record layer SHOULD behave as described below.

If PMTU estimates are available from the underlying transport protocol, they should be made available to upper layer protocols. In particular:

  • For DTLS over UDP, the upper layer protocol SHOULD be allowed to obtain the PMTU estimate maintained in the IP layer.
  • For DTLS over DCCP, the upper layer protocol SHOULD be allowed to obtain the current estimate of the PMTU.
  • For DTLS over TCP or SCTP, which automatically fragment and reassemble datagrams, there is no PMTU limitation. However, the upper layer protocol MUST NOT write any record that exceeds the maximum record size of 2^14 bytes.

Note that DTLS does not defend against spoofed ICMP messages; implementations SHOULD ignore any such messages that indicate PMTUs below the IPv4 and IPv6 minimums of 576 and 1280 bytes respectively

The DTLS record layer SHOULD allow the upper layer protocol to discover the amount of record expansion expected by the DTLS processing.

If there is a transport protocol indication (either via ICMP or via a refusal to send the datagram as in Section 14 of [RFC4340]), then the DTLS record layer MUST inform the upper layer protocol of the error.

The DTLS record layer SHOULD NOT interfere with upper layer protocols performing PMTU discovery, whether via [RFC1191] or [RFC4821] mechanisms. In particular:

  • Where allowed by the underlying transport protocol, the upper layer protocol SHOULD be allowed to set the state of the DF bit (in IPv4) or prohibit local fragmentation (in IPv6).
  • If the underlying transport protocol allows the application to request PMTU probing (e.g., DCCP), the DTLS record layer SHOULD honor this request.

The final issue is the DTLS handshake protocol. From the perspective of the DTLS record layer, this is merely another upper layer protocol. However, DTLS handshakes occur infrequently and involve only a few round trips; therefore, the handshake protocol PMTU handling places a premium on rapid completion over accurate PMTU discovery. In order to allow connections under these circumstances, DTLS implementations SHOULD follow the following rules:

  • If the DTLS record layer informs the DTLS handshake layer that a message is too big, it SHOULD immediately attempt to fragment it, using any existing information about the PMTU.
  • If repeated retransmissions do not result in a response, and the PMTU is unknown, subsequent retransmissions SHOULD back off to a smaller record size, fragmenting the handshake message as appropriate. This standard does not specify an exact number of retransmits to attempt before backing off, but 2-3 seems appropriate.

4.5. Record Payload Protection

Like TLS, DTLS transmits data as a series of protected records. The rest of this section describes the details of that format.

4.5.1. Anti-Replay

Each DTLS record contains a sequence number to provide replay protection. Sequence number verification SHOULD be performed using the following sliding window procedure, borrowed from Section 3.4.3 of [RFC4303].

The received record counter for a session MUST be initialized to zero when that session is established. For each received record, the receiver MUST verify that the record contains a sequence number that does not duplicate the sequence number of any other record received during the lifetime of the session. This check SHOULD happen after deprotecting the record; otherwise the record discard might itself serve as a timing channel for the record number. Note that decompressing the records number is still a potential timing channel for the record number, though a less powerful one than whether it was deprotected.

Duplicates are rejected through the use of a sliding receive window. (How the window is implemented is a local matter, but the following text describes the functionality that the implementation must exhibit.) The receiver SHOULD pick a window large enough to handle any plausible reordering, which depends on the data rate. (The receiver does not notify the sender of the window size.)

The "right" edge of the window represents the highest validated sequence number value received on the session. Records that contain sequence numbers lower than the "left" edge of the window are rejected. Records falling within the window are checked against a list of received records within the window. An efficient means for performing this check, based on the use of a bit mask, is described in Section 3.4.3 of [RFC4303]. If the received record falls within the window and is new, or if the record is to the right of the window, then the record is new.

The window MUST NOT be updated until the record has been deprotected successfully.

4.5.2. Handling Invalid Records

Unlike TLS, DTLS is resilient in the face of invalid records (e.g., invalid formatting, length, MAC, etc.). In general, invalid records SHOULD be silently discarded, thus preserving the association; however, an error MAY be logged for diagnostic purposes. Implementations which choose to generate an alert instead, MUST generate error alerts to avoid attacks where the attacker repeatedly probes the implementation to see how it responds to various types of error. Note that if DTLS is run over UDP, then any implementation which does this will be extremely susceptible to denial-of-service (DoS) attacks because UDP forgery is so easy. Thus, this practice is NOT RECOMMENDED for such transports, both to increase the reliability of DTLS service and to avoid the risk of spoofing attacks sending traffic to unrelated third parties.

If DTLS is being carried over a transport that is resistant to forgery (e.g., SCTP with SCTP-AUTH), then it is safer to send alerts because an attacker will have difficulty forging a datagram that will not be rejected by the transport layer.

4.5.3. AEAD Limits

Section 5.5 of TLS [TLS13] defines limits on the number of records that can be protected using the same keys. These limits are specific to an AEAD algorithm, and apply equally to DTLS. Implementations SHOULD NOT protect more records than allowed by the limit specified for the negotiated AEAD. Implementations SHOULD initiate a key update before reaching this limit.

[TLS13] does not specify a limit for AEAD_AES_128_CCM, but the analysis in Appendix B shows that a limit of 2^23 packets can be used to obtain the same confidentiality protection as the limits specified in TLS.

The usage limits defined in TLS 1.3 exist for protection against attacks on confidentiality and apply to successful applications of AEAD protection. The integrity protections in authenticated encryption also depend on limiting the number of attempts to forge packets. TLS achieves this by closing connections after any record fails an authentication check. In comparison, DTLS ignores any packet that cannot be authenticated, allowing multiple forgery attempts.

Implementations MUST count the number of received packets that fail authentication with each key. If the number of packets that fail authentication exceed a limit that is specific to the AEAD in use, an implementation SHOULD immediately close the connection. Implementations SHOULD initiate a key update with update_requested before reaching this limit. Once a key update has been initiated, the previous keys can be dropped when the limit is reached rather than closing the connection. Applying a limit reduces the probability that an attacker is able to successfully forge a packet; see [AEBounds] and [ROBUST].

For AEAD_AES_128_GCM, AEAD_AES_256_GCM, and AEAD_CHACHA20_POLY1305, the limit on the number of records that fail authentication is 2^36. Note that the analysis in [AEBounds] supports a higher limit for the AEAD_AES_128_GCM and AEAD_AES_256_GCM, but this specification recommends a lower limit. For AEAD_AES_128_CCM, the limit on the number of records that fail authentication is 2^23.5; see Appendix B.

The AEAD_AES_128_CCM_8 AEAD, as used in TLS_AES_128_CCM_SHA256, does not have a limit on the number of records that fail authentication that both limits the probability of forgery by the same amount and does not expose implementations to the risk of denial of service; see Appendix B.3. Therefore, TLS_AES_128_CCM_SHA256 MUST NOT used in DTLS without additional safeguards against forgery. Implementations MUST set usage limits for AEAD_AES_128_CCM_8 based on an understanding of any additional forgery protections that are used.

Any TLS cipher suite that is specified for use with DTLS MUST define limits on the use of the associated AEAD function that preserves margins for both confidentiality and integrity. That is, limits MUST be specified for the number of packets that can be authenticated and for the number packets that can fail authentication. Providing a reference to any analysis upon which values are based - and any assumptions used in that analysis - allows limits to be adapted to varying usage conditions.

5. The DTLS Handshake Protocol

DTLS 1.3 re-uses the TLS 1.3 handshake messages and flows, with the following changes:

  1. To handle message loss, reordering, and fragmentation modifications to the handshake header are necessary.
  2. Retransmission timers are introduced to handle message loss.
  3. A new ACK content type has been added for reliable message delivery of handshake messages.

Note that TLS 1.3 already supports a cookie extension, which is used to prevent denial-of-service attacks. This DoS prevention mechanism is described in more detail below since UDP-based protocols are more vulnerable to amplification attacks than a connection-oriented transport like TCP that performs return-routability checks as part of the connection establishment.

DTLS implementations do not use the TLS 1.3 "compatibility mode" described in Section D.4 of [TLS13]. DTLS servers MUST NOT echo the "session_id" value from the client and endpoints MUST NOT send ChangeCipherSpec messages.

With these exceptions, the DTLS message formats, flows, and logic are the same as those of TLS 1.3.

5.1. Denial-of-Service Countermeasures

Datagram security protocols are extremely susceptible to a variety of DoS attacks. Two attacks are of particular concern:

  1. An attacker can consume excessive resources on the server by transmitting a series of handshake initiation requests, causing the server to allocate state and potentially to perform expensive cryptographic operations.
  2. An attacker can use the server as an amplifier by sending connection initiation messages with a forged source of the victim. The server then sends its response to the victim machine, thus flooding it. Depending on the selected parameters this response message can be quite large, as it is the case for a Certificate message.

In order to counter both of these attacks, DTLS borrows the stateless cookie technique used by Photuris [RFC2522] and IKE [RFC7296]. When the client sends its ClientHello message to the server, the server MAY respond with a HelloRetryRequest message. The HelloRetryRequest message, as well as the cookie extension, is defined in TLS 1.3. The HelloRetryRequest message contains a stateless cookie generated using the technique of [RFC2522]. The client MUST retransmit the ClientHello with the cookie added as an extension. The server then verifies the cookie and proceeds with the handshake only if it is valid. This mechanism forces the attacker/client to be able to receive the cookie, which makes DoS attacks with spoofed IP addresses difficult. This mechanism does not provide any defense against DoS attacks mounted from valid IP addresses.

The DTLS 1.3 specification changes how cookies are exchanged compared to DTLS 1.2. DTLS 1.3 re-uses the HelloRetryRequest message and conveys the cookie to the client via an extension. The client receiving the cookie uses the same extension to place the cookie subsequently into a ClientHello message. DTLS 1.2 on the other hand used a separate message, namely the HelloVerifyRequest, to pass a cookie to the client and did not utilize the extension mechanism. For backwards compatibility reasons, the cookie field in the ClientHello is present in DTLS 1.3 but is ignored by a DTLS 1.3 compliant server implementation.

The exchange is shown in Figure 5. Note that the figure focuses on the cookie exchange; all other extensions are omitted.

The cookie extension is defined in Section 4.2.2 of [TLS13]. When sending the initial ClientHello, the client does not have a cookie yet. In this case, the cookie extension is omitted and the legacy_cookie field in the ClientHello message MUST be set to a zero length vector (i.e., a single zero byte length field).

When responding to a HelloRetryRequest, the client MUST create a new ClientHello message following the description in Section 4.1.2 of [TLS13].

If the HelloRetryRequest message is used, the initial ClientHello and the HelloRetryRequest are included in the calculation of the transcript hash. The computation of the message hash for the HelloRetryRequest is done according to the description in Section 4.4.1 of [TLS13].

The handshake transcript is not reset with the second ClientHello and a stateless server-cookie implementation requires the transcript of the HelloRetryRequest to be stored in the cookie or the internal state of the hash algorithm, since only the hash of the transcript is required for the handshake to complete.

When the second ClientHello is received, the server can verify that the cookie is valid and that the client can receive packets at the given IP address. If the client's apparent IP address is embedded in the cookie, this prevents an attacker from generating an acceptable ClientHello apparently from another user.

One potential attack on this scheme is for the attacker to collect a number of cookies from different addresses where it controls endpoints and then reuse them to attack the server. The server can defend against this attack by changing the secret value frequently, thus invalidating those cookies. If the server wishes to allow legitimate clients to handshake through the transition (e.g., a client received a cookie with Secret 1 and then sent the second ClientHello after the server has changed to Secret 2), the server can have a limited window during which it accepts both secrets. [RFC7296] suggests adding a key identifier to cookies to detect this case. An alternative approach is simply to try verifying with both secrets. It is RECOMMENDED that servers implement a key rotation scheme that allows the server to manage keys with overlapping lifetime.

Alternatively, the server can store timestamps in the cookie and reject cookies that were generated outside a certain interval of time.

DTLS servers SHOULD perform a cookie exchange whenever a new handshake is being performed. If the server is being operated in an environment where amplification is not a problem, the server MAY be configured not to perform a cookie exchange. The default SHOULD be that the exchange is performed, however. In addition, the server MAY choose not to do a cookie exchange when a session is resumed or, more generically, when the DTLS handshake uses a PSK-based key exchange. Clients MUST be prepared to do a cookie exchange with every handshake.

If a server receives a ClientHello with an invalid cookie, it MUST NOT terminate the handshake with an "illegal_parameter" alert. This allows the client to restart the connection from scratch without a cookie.

As described in Section 4.1.4 of [TLS13], clients MUST abort the handshake with an "unexpected_message" alert in response to any second HelloRetryRequest which was sent in the same connection (i.e., where the ClientHello was itself in response to a HelloRetryRequest).

5.2. DTLS Handshake Message Format

In order to support message loss, reordering, and message fragmentation, DTLS modifies the TLS 1.3 handshake header:

    enum {
    } HandshakeType;

    struct {
        HandshakeType msg_type;    /* handshake type */
        uint24 length;             /* bytes in message */
        uint16 message_seq;        /* DTLS-required field */
        uint24 fragment_offset;    /* DTLS-required field */
        uint24 fragment_length;    /* DTLS-required field */
        select (HandshakeType) {
            case client_hello:          ClientHello;
            case server_hello:          ServerHello;
            case end_of_early_data:     EndOfEarlyData;
            case encrypted_extensions:  EncryptedExtensions;
            case certificate_request:   CertificateRequest;
            case certificate:           Certificate;
            case certificate_verify:    CertificateVerify;
            case finished:              Finished;
            case new_session_ticket:    NewSessionTicket;
            case key_update:            KeyUpdate;
        } body;
    } Handshake;

The first message each side transmits in each association always has message_seq = 0. Whenever a new message is generated, the message_seq value is incremented by one. When a message is retransmitted, the old message_seq value is re-used, i.e., not incremented. From the perspective of the DTLS record layer, the retransmission is a new record. This record will have a new DTLSPlaintext.sequence_number value.

Note: In DTLS 1.2 the message_seq was reset to zero in case of a rehandshake (i.e., renegotiation). On the surface, a rehandshake in DTLS 1.2 shares similarities with a post-handshake message exchange in DTLS 1.3. However, in DTLS 1.3 the message_seq is not reset to allow distinguishing a retransmission from a previously sent post-handshake message from a newly sent post-handshake message.

DTLS implementations maintain (at least notionally) a next_receive_seq counter. This counter is initially set to zero. When a handshake message is received, if its message_seq value matches next_receive_seq, next_receive_seq is incremented and the message is processed. If the sequence number is less than next_receive_seq, the message MUST be discarded. If the sequence number is greater than next_receive_seq, the implementation SHOULD queue the message but MAY discard it. (This is a simple space/bandwidth tradeoff).

In addition to the handshake messages that are deprecated by the TLS 1.3 specification, DTLS 1.3 furthermore deprecates the HelloVerifyRequest message originally defined in DTLS 1.0. DTLS 1.3-compliant implements MUST NOT use the HelloVerifyRequest to execute a return-routability check. A dual-stack DTLS 1.2/DTLS 1.3 client MUST, however, be prepared to interact with a DTLS 1.2 server.

5.3. ClientHello Message

The format of the ClientHello used by a DTLS 1.3 client differs from the TLS 1.3 ClientHello format as shown below.

    uint16 ProtocolVersion;
    opaque Random[32];

    uint8 CipherSuite[2];    /* Cryptographic suite selector */

    struct {
        ProtocolVersion legacy_version = { 254,253 }; // DTLSv1.2
        Random random;
        opaque legacy_session_id<0..32>;
        opaque legacy_cookie<0..2^8-1>;                  // DTLS
        CipherSuite cipher_suites<2..2^16-2>;
        opaque legacy_compression_methods<1..2^8-1>;
        Extension extensions<8..2^16-1>;
    } ClientHello;
In previous versions of DTLS, this field was used for version negotiation and represented the highest version number supported by the client. Experience has shown that many servers do not properly implement version negotiation, leading to "version intolerance" in which the server rejects an otherwise acceptable ClientHello with a version number higher than it supports. In DTLS 1.3, the client indicates its version preferences in the "supported_versions" extension (see Section 4.2.1 of [TLS13]) and the legacy_version field MUST be set to {254, 253}, which was the version number for DTLS 1.2. The version fields for DTLS 1.0 and DTLS 1.2 are 0xfeff and 0xfefd (to match the wire versions) but the version field for DTLS 1.3 is 0x0304.
Same as for TLS 1.3.
Same as for TLS 1.3.
A DTLS 1.3-only client MUST set the legacy_cookie field to zero length. If a DTLS 1.3 ClientHello is received with any other value in this field, the server MUST abort the handshake with an "illegal_parameter" alert.
Same as for TLS 1.3.
Same as for TLS 1.3.
Same as for TLS 1.3.

5.4. Handshake Message Fragmentation and Reassembly

Each DTLS message MUST fit within a single transport layer datagram. However, handshake messages are potentially bigger than the maximum record size. Therefore, DTLS provides a mechanism for fragmenting a handshake message over a number of records, each of which can be transmitted separately, thus avoiding IP fragmentation.

When transmitting the handshake message, the sender divides the message into a series of N contiguous data ranges. The ranges MUST NOT overlap. The sender then creates N handshake messages, all with the same message_seq value as the original handshake message. Each new message is labeled with the fragment_offset (the number of bytes contained in previous fragments) and the fragment_length (the length of this fragment). The length field in all messages is the same as the length field of the original message. An unfragmented message is a degenerate case with fragment_offset=0 and fragment_length=length. Each range MUST be delivered in a single UDP datagram.

When a DTLS implementation receives a handshake message fragment, it MUST buffer it until it has the entire handshake message. DTLS implementations MUST be able to handle overlapping fragment ranges. This allows senders to retransmit handshake messages with smaller fragment sizes if the PMTU estimate changes.

Note that as with TLS, multiple handshake messages may be placed in the same DTLS record, provided that there is room and that they are part of the same flight. Thus, there are two acceptable ways to pack two DTLS messages into the same datagram: in the same record or in separate records.

5.5. End Of Early Data

The DTLS 1.3 handshake has one important difference from the TLS 1.3 handshake: the EndOfEarlyData message is omitted both from the wire and the handshake transcript: because DTLS records have epochs, EndOfEarlyData is not necessary to determine when the early data is complete, and because DTLS is lossy, attackers can trivially mount the deletion attacks that EndOfEarlyData prevents in TLS. Servers SHOULD aggressively age out the epoch 1 keys upon receiving the first epoch 2 record and SHOULD NOT accept epoch 1 data after the first epoch 3 record is received. (See Section 6.1 for the definitions of each epoch.)

5.6. DTLS Handshake Flights

DTLS messages are grouped into a series of message flights, according to the diagrams below.

Client                                             Server

ClientHello                                                 +----------+
 + key_share*                                               | Flight 1 |
 + pre_shared_key*      -------->                           +----------+

                        <--------        HelloRetryRequest  | Flight 2 |
                                          + cookie          +----------+

ClientHello                                                 +----------+
 + key_share*                                               | Flight 3 |
 + pre_shared_key*      -------->                           +----------+
 + cookie

                                              + key_share*
                                         + pre_shared_key*  +----------+
                                     {EncryptedExtensions}  | Flight 4 |
                                     {CertificateRequest*}  +----------+
                        <--------               {Finished}
                                       [Application Data*]

 {Certificate*}                                             +----------+
 {CertificateVerify*}                                       | Flight 5 |
 {Finished}             -------->                           +----------+
 [Application Data]

                        <--------                    [ACK]  | Flight 6 |
                                       [Application Data*]  +----------+

 [Application Data]     <------->      [Application Data]
Figure 6: Message flights for a full DTLS Handshake (with cookie exchange)
 ClientHello                                              +----------+
  + pre_shared_key                                        | Flight 1 |
  + key_share*         -------->                          +----------+

                                        + pre_shared_key  +----------+
                                            + key_share*  | Flight 2 |
                                   {EncryptedExtensions}  +----------+
                       <--------              {Finished}
                                     [Application Data*]
 {Finished}            -------->                          | Flight 3 |
 [Application Data*]                                      +----------+

                       <--------                   [ACK]  | Flight 4 |
                                     [Application Data*]  +----------+

 [Application Data]    <------->      [Application Data]
Figure 7: Message flights for resumption and PSK handshake (without cookie exchange)
Client                                            Server

  + early_data
  + psk_key_exchange_modes                                +----------+
  + key_share*                                            | Flight 1 |
  + pre_shared_key                                        +----------+
 (Application Data*)     -------->

                                        + pre_shared_key
                                            + key_share*  +----------+
                                   {EncryptedExtensions}  | Flight 2 |
                                              {Finished}  +----------+
                       <--------     [Application Data*]

 {Finished}            -------->                          | Flight 3 |
 [Application Data*]                                      +----------+

                       <--------                   [ACK]  | Flight 4 |
                                     [Application Data*]  +----------+

 [Application Data]    <------->      [Application Data]
Figure 8: Message flights for the Zero-RTT handshake
Client                                            Server

                       <--------       [NewSessionTicket] | Flight 1 |

[ACK]                  -------->                          | Flight 2 |
Figure 9: Message flights for the new session ticket message

Note: The application data sent by the client is not included in the timeout and retransmission calculation.

5.7. Timeout and Retransmission

5.7.1. State Machine

DTLS uses a simple timeout and retransmission scheme with the state machine shown in Figure 10. Because DTLS clients send the first message (ClientHello), they start in the PREPARING state. DTLS servers start in the WAITING state, but with empty buffers and no retransmit timer.

                             | PREPARING |
                +----------> |           |
                |            |           |
                |            +-----------+
                |                  |
                |                  | Buffer next flight
                |                  |
                |                 \|/
                |            +-----------+
                |            |           |
                |            |  SENDING  |<------------------+
                |            |           |                   |
                |            +-----------+                   |
        Receive |                  |                         |
           next |                  | Send flight or partial  |
         flight |                  | flight                  |
                |                  |                         |
                |                  | Set retransmit timer    |
                |                 \|/                        |
                |            +-----------+                   |
                |            |           |                   |
                +------------|  WAITING  |-------------------+
                |     +----->|           |   Timer expires   |
                |     |      +-----------+                   |
                |     |          |  |   |                    |
                |     |          |  |   |                    |
                |     +----------+  |   +--------------------+
                |    Receive record |   Read retransmit or ACK
        Receive |     Send ACK      |
           last |                   |
         flight |                   | Receive ACK
                |                   | for last flight
               \|/                  |
            +-----------+           |
            |           | <---------+
            | FINISHED  |
            |           |
                |  /|\
                |   |
                |   |

          Server read retransmit
              Retransmit ACK
Figure 10: DTLS timeout and retransmission state machine

The state machine has four basic states: PREPARING, SENDING, WAITING, and FINISHED.

In the PREPARING state, the implementation does whatever computations are necessary to prepare the next flight of messages. It then buffers them up for transmission (emptying the buffer first) and enters the SENDING state.

In the SENDING state, the implementation transmits the buffered flight of messages. If the implementation has received one or more ACKs (see Section 7) from the peer, then it SHOULD omit any messages or message fragments which have already been ACKed. Once the messages have been sent, the implementation then sets a retransmit timer and enters the WAITING state.

There are four ways to exit the WAITING state:

  1. The retransmit timer expires: the implementation transitions to the SENDING state, where it retransmits the flight, resets the retransmit timer, and returns to the WAITING state.
  2. The implementation reads an ACK from the peer: upon receiving an ACK for a partial flight (as mentioned in Section 7.1), the implementation transitions to the SENDING state, where it retransmits the unacked portion of the flight, resets the retransmit timer, and returns to the WAITING state. Upon receiving an ACK for a complete flight, the implementation cancels all retransmissions and either remains in WAITING, or, if the ACK was for the final flight, transitions to FINISHED.
  3. The implementation reads a retransmitted flight from the peer: the implementation transitions to the SENDING state, where it retransmits the flight, resets the retransmit timer, and returns to the WAITING state. The rationale here is that the receipt of a duplicate message is the likely result of timer expiry on the peer and therefore suggests that part of one's previous flight was lost.
  4. The implementation receives some or all next flight of messages: if this is the final flight of messages, the implementation transitions to FINISHED. If the implementation needs to send a new flight, it transitions to the PREPARING state. Partial reads (whether partial messages or only some of the messages in the flight) may also trigger the implementation to send an ACK, as described in Section 7.1.

Because DTLS clients send the first message (ClientHello), they start in the PREPARING state. DTLS servers start in the WAITING state, but with empty buffers and no retransmit timer.

In addition, for at least twice the default MSL defined for [RFC0793], when in the FINISHED state, the server MUST respond to retransmission of the client's second flight with a retransmit of its ACK.

Note that because of packet loss, it is possible for one side to be sending application data even though the other side has not received the first side's Finished message. Implementations MUST either discard or buffer all application data records for the new epoch until they have received the Finished message for that epoch. Implementations MAY treat receipt of application data with a new epoch prior to receipt of the corresponding Finished message as evidence of reordering or packet loss and retransmit their final flight immediately, shortcutting the retransmission timer.

5.7.2. Timer Values

Though timer values are the choice of the implementation, mishandling of the timer can lead to serious congestion problems; for example, if many instances of a DTLS time out early and retransmit too quickly on a congested link. Implementations SHOULD use an initial timer value of 100 msec (the minimum defined in RFC 6298 [RFC6298]) and double the value at each retransmission, up to no less than the RFC 6298 maximum of 60 seconds. Application specific profiles, such as those used for the Internet of Things environment, may recommend longer timer values. Note that a 100 msec timer is recommended rather than the 3-second RFC 6298 default in order to improve latency for time-sensitive applications. Because DTLS only uses retransmission for handshake and not dataflow, the effect on congestion should be minimal.

Implementations SHOULD retain the current timer value until a transmission without loss occurs, at which time the value may be reset to the initial value. After a long period of idleness, no less than 10 times the current timer value, implementations may reset the timer to the initial value.

5.7.3. State machine duplication for post-handshake messages

DTLS 1.3 makes use of the following categories of post-handshake messages:

  1. NewSessionTicket
  2. KeyUpdate
  3. NewConnectionId
  4. RequestConnectionId
  5. Post-handshake client authentication

Messages of each category can be sent independently, and reliability is established via independent state machines each of which behaves as described in Section 5.7.1. For example, if a server sends a NewSessionTicket and a CertificateRequest message, two independent state machines will be created.

As explained in the corresponding sections, sending multiple instances of messages of a given category without having completed earlier transmissions is allowed for some categories, but not for others. Specifically, a server MAY send multiple NewSessionTicket messages at once without awaiting ACKs for earlier NewSessionTicket first. Likewise, a server MAY send multiple CertificateRequest messages at once without having completed earlier client authentication requests before. In contrast, implementations MUST NOT have send KeyUpdate, NewConnectionId or RequestConnectionId message if an earlier message of the same type has not yet been acknowledged.

Note: Except for post-handshake client authentication, which involves handshake messages in both directions, post-handshake messages are single-flight, and their respective state machines on the sender side reduce to waiting for an ACK and retransmitting the original message. In particular, note that a RequestConnectionId message does not force the receiver to send a NewConnectionId message in reply, and both messages are therefore treated independently.

Creating and correctly updating multiple state machines requires feedback from the handshake logic to the state machine layer, indicating which message belongs to which state machine. For example, if a server sends multiple CertificateRequest messages and receives a Certificate message in response, the corresponding state machine can only be determined after inspecting the certificate_request_context field. Similarly, a server sending a single CertificateRequest and receiving a NewConnectionId message in response can only decide that the NewConnectionId message should be treated through an independent state machine after inspecting the handshake message type.

5.8. CertificateVerify and Finished Messages

CertificateVerify and Finished messages have the same format as in TLS 1.3. Hash calculations include entire handshake messages, including DTLS-specific fields: message_seq, fragment_offset, and fragment_length. However, in order to remove sensitivity to handshake message fragmentation, the CertificateVerify and the Finished messages MUST be computed as if each handshake message had been sent as a single fragment following the algorithm described in Section 4.4.3 and Section 4.4.4 of [TLS13], respectively.

5.9. Cryptographic Label Prefix

Section 7.1 of [TLS13] specifies that HKDF-Expand-Label uses a label prefix of "tls13 ". For DTLS 1.3, that label SHALL be "dtls13". This ensures key separation between DTLS 1.3 and TLS 1.3. Note that there is no trailing space; this is necessary in order to keep the overall label size inside of one hash iteration because "DTLS" is one letter longer than "TLS".

5.10. Alert Messages

Note that Alert messages are not retransmitted at all, even when they occur in the context of a handshake. However, a DTLS implementation which would ordinarily issue an alert SHOULD generate a new alert message if the offending record is received again (e.g., as a retransmitted handshake message). Implementations SHOULD detect when a peer is persistently sending bad messages and terminate the local connection state after such misbehavior is detected.

5.11. Establishing New Associations with Existing Parameters

If a DTLS client-server pair is configured in such a way that repeated connections happen on the same host/port quartet, then it is possible that a client will silently abandon one connection and then initiate another with the same parameters (e.g., after a reboot). This will appear to the server as a new handshake with epoch=0. In cases where a server believes it has an existing association on a given host/port quartet and it receives an epoch=0 ClientHello, it SHOULD proceed with a new handshake but MUST NOT destroy the existing association until the client has demonstrated reachability either by completing a cookie exchange or by completing a complete handshake including delivering a verifiable Finished message. After a correct Finished message is received, the server MUST abandon the previous association to avoid confusion between two valid associations with overlapping epochs. The reachability requirement prevents off-path/blind attackers from destroying associations merely by sending forged ClientHellos.

Note: it is not always possible to distinguish which association a given record is from. For instance, if the client performs a handshake, abandons the connection, and then immediately starts a new handshake, it may not be possible to tell which connection a given protected record is for. In these cases, trial decryption MAY be necessary, though implementations could use CIDs.

6. Example of Handshake with Timeout and Retransmission

The following is an example of a handshake with lost packets and retransmissions.

Client                                                Server
------                                                ------

 Record 0                  -------->

                             X<-----                 Record 0
                             (lost)               ServerHello

                           <--------                 Record 1

 Record 1                  -------->
 ACK []

                           <--------                 Record 2

 Record 2                  -------->

                           <--------               Record 3
                                                    ACK [2]

Figure 11: Example DTLS exchange illustrating message loss

6.1. Epoch Values and Rekeying

A recipient of a DTLS message needs to select the correct keying material in order to process an incoming message. With the possibility of message loss and re-ordering, an identifier is needed to determine which cipher state has been used to protect the record payload. The epoch value fulfills this role in DTLS. In addition to the TLS 1.3-defined key derivation steps, see Section 7 of [TLS13], a sender may want to rekey at any time during the lifetime of the connection. It therefore needs to indicate that it is updating its sending cryptographic keys.

This version of DTLS assigns dedicated epoch values to messages in the protocol exchange to allow identification of the correct cipher state:

  • epoch value (0) is used with unencrypted messages. There are three unencrypted messages in DTLS, namely ClientHello, ServerHello, and HelloRetryRequest.
  • epoch value (1) is used for messages protected using keys derived from client_early_traffic_secret. Note this epoch is skipped if the client does not offer early data.
  • epoch value (2) is used for messages protected using keys derived from [sender]_handshake_traffic_secret. Messages transmitted during the initial handshake, such as EncryptedExtensions, CertificateRequest, Certificate, CertificateVerify, and Finished belong to this category. Note, however, post-handshake are protected under the appropriate application traffic key and are not included in this category.
  • epoch value (3) is used for payloads protected using keys derived from the initial [sender]_application_traffic_secret_0. This may include handshake messages, such as post-handshake messages (e.g., a NewSessionTicket message).
  • epoch value (4 to 2^16-1) is used for payloads protected using keys from the [sender]_application_traffic_secret_N (N>0).

Using these reserved epoch values a receiver knows what cipher state has been used to encrypt and integrity protect a message. Implementations that receive a payload with an epoch value for which no corresponding cipher state can be determined MUST generate a "unexpected_message" alert. For example, if a client incorrectly uses epoch value 5 when sending early application data in a 0-RTT exchange. A server will not be able to compute the appropriate keys and will therefore have to respond with an alert.

Note that epoch values do not wrap. If a DTLS implementation would need to wrap the epoch value, it MUST terminate the connection.

The traffic key calculation is described in Section 7.3 of [TLS13].

Figure 12 illustrates the epoch values in an example DTLS handshake.

Client                                             Server
------                                             ------


                            <--------             ServerHello

 ClientHello                -------->

                            <--------             ServerHello

 {Certificate}              -------->

                            <--------                   [ACK]

 [Application Data]         -------->

                            <--------      [Application Data]

                         Some time later ...
                 (Post-Handshake Message Exchange)

                            <--------      [NewSessionTicket]

 [ACK]                      -------->

                         Some time later ...

                            <--------      [Application Data]
 [Application Data]         -------->
Figure 12: Example DTLS exchange with epoch information

7. ACK Message

The ACK message is used by an endpoint to indicate which handshake records it has received and processed from the other side. ACK is not a handshake message but is rather a separate content type, with code point TBD (proposed, 25). This avoids having ACK being added to the handshake transcript. Note that ACKs can still be sent in the same UDP datagram as handshake records.

    struct {
        RecordNumber record_numbers<0..2^16-1>;
    } ACK;
a list of the records containing handshake messages in the current flight which the endpoint has received and either processed or buffered, in numerically increasing order.

Implementations MUST NOT acknowledge records containing handshake messages or fragments which have not been processed or buffered. Otherwise, deadlock can ensue. As an example, implementations MUST NOT send ACKs for handshake messages which they discard because they are not the next expected message.

During the handshake, ACKs only cover the current outstanding flight (this is possible because DTLS is generally a lockstep protocol). Thus, an ACK from the server would not cover both the ClientHello and the client's Certificate. Implementations can accomplish this by clearing their ACK list upon receiving the start of the next flight.

After the handshake, ACKs SHOULD be sent once for each received and processed handshake record (potentially subject to some delay) and MAY cover more than one flight. This includes messages which are discarded because a previous copy has been received.

During the handshake, ACK records MUST be sent with an epoch that is equal to or higher than the record which is being acknowledged. Note that some care is required when processing flights spanning multiple epochs. For instance, if the client receives only the Server Hello and Certificate and wishes to ACK them in a single record, it must do so in epoch 2, as it is required to use an epoch greater than or equal to 2 and cannot yet send with any greater epoch. Implementations SHOULD simply use the highest current sending epoch, which will generally be the highest available. After the handshake, implementations MUST use the highest available sending epoch.

7.1. Sending ACKs

When an implementation detects a disruption in the receipt of the current incoming flight, it SHOULD generate an ACK that covers the messages from that flight which it has received and processed so far. Implementations have some discretion about which events to treat as signs of disruption, but it is RECOMMENDED that they generate ACKs under two circumstances:

  • When they receive a message or fragment which is out of order, either because it is not the next expected message or because it is not the next piece of the current message.
  • When they have received part of a flight and do not immediately receive the rest of the flight (which may be in the same UDP datagram). A reasonable approach here is to set a timer for 1/4 the current retransmit timer value when the first record in the flight is received and then send an ACK when that timer expires.

In general, flights MUST be ACKed unless they are implicitly acknowledged. In the present specification the following flights are implicitly acknowledged by the receipt of the next flight, which generally immediately follows the flight,

  1. Handshake flights other than the client's final flight
  2. The server's post-handshake CertificateRequest.

ACKs SHOULD NOT be sent for these flights unless generating the responding flight takes significant time. In this case, implementations MAY send explicit ACKs for the complete received flight even though it will eventually also be implicitly acknowledged through the responding flight. A notable example for this is the case of post-handshake client authentication in constrained environments, where generating the CertificateVerify message can take considerable time on the client. All other flights MUST be ACKed. Implementations MAY acknowledge the records corresponding to each transmission of each flight or simply acknowledge the most recent one. In general, implementations SHOULD ACK as many received packets as can fit into the ACK record, as this provides the most complete information and thus reduces the chance of spurious retransmission; if space is limited, implementations SHOULD favor including records which have not yet been acknowledged.

Note: While some post-handshake messages follow a request/response pattern, this does not necessarily imply receipt. For example, a KeyUpdate sent in response to a KeyUpdate with update_requested does not implicitly acknowledge that message because the KeyUpdates might have crossed in flight.

ACKs MUST NOT be sent for other records of any content type other than handshake or for records which cannot be unprotected.

Note that in some cases it may be necessary to send an ACK which does not contain any record numbers. For instance, a client might receive an EncryptedExtensions message prior to receiving a ServerHello. Because it cannot decrypt the EncryptedExtensions, it cannot safely acknowledge it (as it might be damaged). If the client does not send an ACK, the server will eventually retransmit its first flight, but this might take far longer than the actual round trip time between client and server. Having the client send an empty ACK shortcuts this process.

7.2. Receiving ACKs

When an implementation receives an ACK, it SHOULD record that the messages or message fragments sent in the records being ACKed were received and omit them from any future retransmissions. Upon receipt of an ACK that leaves it with only some messages from a flight having been acknowledged an implementation SHOULD retransmit the unacknowledged messages or fragments. Note that this requires implementations to track which messages appear in which records. Once all the messages in a flight have been acknowledged, the implementation MUST cancel all retransmissions of that flight. Implementations MUST treat a record as having been acknowledged if it appears in any ACK; this prevents spurious retransmission in cases where a flight is very large and the receiver is forced to elide acknowledgements for records which have already been ACKed. As noted above, the receipt of any record responding to a given flight MUST be taken as an implicit acknowledgement for the entire flight.

7.3. Design Rational

ACK messages are used in two circumstances, namely :

  • on sign of disruption, or lack of progress, and
  • to indicate complete receipt of the last flight in a handshake.

In the first case the use of the ACK message is optional because the peer will retransmit in any case and therefore the ACK just allows for selective retransmission, as opposed to the whole flight retransmission in previous versions of DTLS. For instance in the flow shown in Figure 11 if the client does not send the ACK message when it received and processed record 1 indicating loss of record 0, the entire flight would be retransmitted. When DTLS 1.3 is used in deployments with loss networks, such as low-power, long range radio networks as well as low-power mesh networks, the use of ACKs is recommended.

The use of the ACK for the second case is mandatory for the proper functioning of the protocol. For instance, the ACK message sent by the client in Figure 12, acknowledges receipt and processing of record 2 (containing the NewSessionTicket message) and if it is not sent the server will continue retransmission of the NewSessionTicket indefinitely.

8. Key Updates

As with TLS 1.3, DTLS 1.3 implementations send a KeyUpdate message to indicate that they are updating their sending keys. As with other handshake messages with no built-in response, KeyUpdates MUST be acknowledged. In order to facilitate epoch reconstruction Section 4.2.2 implementations MUST NOT send with the new keys or send a new KeyUpdate until the previous KeyUpdate has been acknowledged (this avoids having too many epochs in active use).

Due to loss and/or re-ordering, DTLS 1.3 implementations may receive a record with an older epoch than the current one (the requirements above preclude receiving a newer record). They SHOULD attempt to process those records with that epoch (see Section 4.2.2 for information on determining the correct epoch), but MAY opt to discard such out-of-epoch records.

Due to the possibility of an ACK message for a KeyUpdate being lost and thereby preventing the sender of the KeyUpdate from updating its keying material, receivers MUST retain the pre-update keying material until receipt and successful decryption of a message using the new keys.

9. Connection ID Updates

If the client and server have negotiated the "connection_id" extension [I-D.ietf-tls-dtls-connection-id], either side can send a new CID which it wishes the other side to use in a NewConnectionId message.

    enum {
        cid_immediate(0), cid_spare(1), (255)
    } ConnectionIdUsage;

    opaque ConnectionId<0..2^8-1>;

    struct {
        ConnectionIds cids<0..2^16-1>;
        ConnectionIdUsage usage;
    } NewConnectionId;
Indicates the set of CIDs which the sender wishes the peer to use.
Indicates whether the new CIDs should be used immediately or are spare. If usage is set to "cid_immediate", then one of the new CID MUST be used immediately for all future records. If it is set to "cid_spare", then either existing or new CID MAY be used.

Endpoints SHOULD use receiver-provided CIDs in the order they were provided. Endpoints MUST NOT have more than one NewConnectionId message outstanding.

If the client and server have negotiated the "connection_id" extension, either side can request a new CID using the RequestConnectionId message.

    struct {
      uint8 num_cids;
    } RequestConnectionId;
The number of CIDs desired.

Endpoints SHOULD respond to RequestConnectionId by sending a NewConnectionId with usage "cid_spare" containing num_cid CIDs soon as possible. Endpoints MUST NOT send a RequestConnectionId message when an existing request is still unfulfilled; this implies that endpoints needs to request new CIDs well in advance. An endpoint MAY ignore requests, which it considers excessive (though they MUST be acknowledged as usual).

Endpoints MUST NOT send either of these messages if they did not negotiate a CID. If an implementation receives these messages when CIDs were not negotiated, it MUST abort the connection with an unexpected_message alert.

9.1. Connection ID Example

Below is an example exchange for DTLS 1.3 using a single CID in each direction.

Note: The connection_id extension is defined in [I-D.ietf-tls-dtls-connection-id], which is used in ClientHello and ServerHello messages.

Client                                             Server
------                                             ------


                            <--------       HelloRetryRequest

ClientHello                 -------->

                            <--------             ServerHello

Certificate                -------->
                           <--------                      Ack

Application Data           ========>
                           <========         Application Data
Figure 13: Example DTLS 1.3 Exchange with CIDs

If no CID is negotiated, then the receiver MUST reject any records it receives that contain a CID.

10. Application Data Protocol

Application data messages are carried by the record layer and are fragmented and encrypted based on the current connection state. The messages are treated as transparent data to the record layer.

11. Security Considerations

Security issues are discussed primarily in [TLS13].

The primary additional security consideration raised by DTLS is that of denial of service. DTLS includes a cookie exchange designed to protect against denial of service. However, implementations that do not use this cookie exchange are still vulnerable to DoS. In particular, DTLS servers that do not use the cookie exchange may be used as attack amplifiers even if they themselves are not experiencing DoS. Therefore, DTLS servers SHOULD use the cookie exchange unless there is good reason to believe that amplification is not a threat in their environment. Clients MUST be prepared to do a cookie exchange with every handshake.

DTLS implementations MUST NOT update their sending address in response to packets from a different address unless they first perform some reachability test; no such test is defined in this specification. Even with such a test, an on-path adversary can also black-hole traffic or create a reflection attack against third parties because a DTLS peer has no means to distinguish a genuine address update event (for example, due to a NAT rebinding) from one that is malicious. This attack is of concern when there is a large asymmetry of request/response message sizes.

With the exception of order protection and non-replayability, the security guarantees for DTLS 1.3 are the same as TLS 1.3. While TLS always provides order protection and non-replayability, DTLS does not provide order protection and may not provide replay protection.

Unlike TLS implementations, DTLS implementations SHOULD NOT respond to invalid records by terminating the connection.

If implementations process out-of-epoch records as recommended in Section 8, then this creates a denial of service risk since an adversary could inject records with fake epoch values, forcing the recipient to compute the next-generation application_traffic_secret using the HKDF-Expand-Label construct to only find out that the message was does not pass the AEAD cipher processing. The impact of this attack is small since the HKDF-Expand-Label only performs symmetric key hashing operations. Implementations which are concerned about this form of attack can discard out-of-epoch records.

The security and privacy properties of the CID for DTLS 1.3 builds on top of what is described in [I-D.ietf-tls-dtls-connection-id]. There are, however, several improvements:

12. Changes to DTLS 1.2

Since TLS 1.3 introduces a large number of changes to TLS 1.2, the list of changes from DTLS 1.2 to DTLS 1.3 is equally large. For this reason this section focuses on the most important changes only.

13. IANA Considerations

IANA is requested to allocate a new value in the "TLS ContentType" registry for the ACK message, defined in Section 7, with content type 26. The value for the "DTLS-OK" column is "Y". IANA is requested to reserve the content type range 32-63 so that content types in this range are not allocated.

IANA is requested to allocate two values in the "TLS Handshake Type" registry, defined in [TLS13], for RequestConnectionId (TBD), and NewConnectionId (TBD), as defined in this document. The value for the "DTLS-OK" columns are "Y".

14. References

14.1. Normative References

Nir, Y. and A. Langley, "ChaCha20 and Poly1305 for IETF Protocols", RFC 8439, DOI 10.17487/RFC8439, , <>.
Rescorla, E., Tschofenig, H., and T. Fossati, "Connection Identifiers for DTLS 1.2", Work in Progress, Internet-Draft, draft-ietf-tls-dtls-connection-id-07, , <>.
Postel, J., "User Datagram Protocol", STD 6, RFC 768, DOI 10.17487/RFC0768, , <>.
Postel, J., "Transmission Control Protocol", STD 7, RFC 793, DOI 10.17487/RFC0793, , <>.
Mogul, J. and S. Deering, "Path MTU discovery", RFC 1191, DOI 10.17487/RFC1191, , <>.
Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", BCP 14, RFC 2119, DOI 10.17487/RFC2119, , <>.
Conta, A., Deering, S., and M. Gupta, Ed., "Internet Control Message Protocol (ICMPv6) for the Internet Protocol Version 6 (IPv6) Specification", STD 89, RFC 4443, DOI 10.17487/RFC4443, , <>.
Mathis, M. and J. Heffner, "Packetization Layer Path MTU Discovery", RFC 4821, DOI 10.17487/RFC4821, , <>.
Paxson, V., Allman, M., Chu, J., and M. Sargent, "Computing TCP's Retransmission Timer", RFC 6298, DOI 10.17487/RFC6298, , <>.
Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC 2119 Key Words", BCP 14, RFC 8174, DOI 10.17487/RFC8174, , <>.
Rescorla, E., "The Transport Layer Security (TLS) Protocol Version 1.3", RFC 8446, DOI 10.17487/RFC8446, , <>.

14.2. Informative References

Luykx, A. and K. Paterson, "Limits on Authenticated Encryption Use in TLS", , <>.
Jonsson, J., "On the Security of CTR + CBC-MAC", Selected Areas in Cryptography pp. 76-93, DOI 10.1007/3-540-36492-7_7, , <>.
Karn, P. and W. Simpson, "Photuris: Session-Key Management Protocol", RFC 2522, DOI 10.17487/RFC2522, , <>.
Kent, S., "IP Encapsulating Security Payload (ESP)", RFC 4303, DOI 10.17487/RFC4303, , <>.
Kohler, E., Handley, M., and S. Floyd, "Datagram Congestion Control Protocol (DCCP)", RFC 4340, DOI 10.17487/RFC4340, , <>.
Dierks, T. and E. Rescorla, "The Transport Layer Security (TLS) Protocol Version 1.1", RFC 4346, DOI 10.17487/RFC4346, , <>.
Rescorla, E. and N. Modadugu, "Datagram Transport Layer Security", RFC 4347, DOI 10.17487/RFC4347, , <>.
Stewart, R., Ed., "Stream Control Transmission Protocol", RFC 4960, DOI 10.17487/RFC4960, , <>.
Phelan, T., "Datagram Transport Layer Security (DTLS) over the Datagram Congestion Control Protocol (DCCP)", RFC 5238, DOI 10.17487/RFC5238, , <>.
Dierks, T. and E. Rescorla, "The Transport Layer Security (TLS) Protocol Version 1.2", RFC 5246, DOI 10.17487/RFC5246, , <>.
Rescorla, E. and N. Modadugu, "Datagram Transport Layer Security Version 1.2", RFC 6347, DOI 10.17487/RFC6347, , <>.
Kaufman, C., Hoffman, P., Nir, Y., Eronen, P., and T. Kivinen, "Internet Key Exchange Protocol Version 2 (IKEv2)", STD 79, RFC 7296, DOI 10.17487/RFC7296, , <>.
Sheffer, Y., Holz, R., and P. Saint-Andre, "Recommendations for Secure Use of Transport Layer Security (TLS) and Datagram Transport Layer Security (DTLS)", BCP 195, RFC 7525, DOI 10.17487/RFC7525, , <>.
Fischlin, M., Günther, F., and C. Janson, "Robust Channels: Handling Unreliable Networks in the Record Layers of QUIC and DTLS 1.3", , <>.

Appendix A. Protocol Data Structures and Constant Values

This section provides the normative protocol types and constants definitions.

%%## Record Layer %%## Handshake Protocol %%## ACKs %%## Connection ID Management

Appendix B. Analysis of Limits on CCM Usage

TLS [TLS13] and [AEBounds] do not specify limits on key usage for AEAD_AES_128_CCM. However, any AEAD that is used with DTLS requires limits on use that ensure that both confidentiality and integrity are preserved. This section documents that analysis for AEAD_AES_128_CCM.

[CCM-ANALYSIS] is used as the basis of this analysis. The results of that analysis are used to derive usage limits that are based on those chosen in [TLS13].

This analysis uses symbols for multiplication (*), division (/), and exponentiation (^), plus parentheses for establishing precedence. The following symbols are also used:

The size of the authentication tag in bits. For this cipher, t is 128.
The size of the block function in bits. For this cipher, n is 128.
The number of blocks in each packet (see below).
The number of genuine packets created and protected by endpoints. This value is the bound on the number of packets that can be protected before updating keys.
The number of forged packets that endpoints will accept. This value is the bound on the number of forged packets that an endpoint can reject before updating keys.

The analysis of AEAD_AES_128_CCM relies on a count of the number of block operations involved in producing each message. For simplicity, and to match the analysis of other AEAD functions in [AEBounds], this analysis assumes a packet length of 2^10 blocks and a packet size limit of 2^14.

For AEAD_AES_128_CCM, the total number of block cipher operations is the sum of: the length of the associated data in blocks, the length of the ciphertext in blocks, the length of the plaintext in blocks, plus 1. In this analysis, this is simplified to a value of twice the maximum length of a record in blocks (that is, 2l = 2^11). This simplification is based on the associated data being limited to one block.

B.1. Confidentiality Limits

For confidentiality, Theorem 2 in [CCM-ANALYSIS] establishes that an attacker gains a distinguishing advantage over an ideal pseudorandom permutation (PRP) of no more than:

(2l * q)^2 / 2^n

For a target advantage of 2^-60, which matches that used by [TLS13], this results in the relation:

q <= 2^23

That is, endpoints cannot protect more than 2^23 packets with the same set of keys without causing an attacker to gain an larger advantage than the target of 2^-60.

B.2. Integrity Limits

For integrity, Theorem 1 in [CCM-ANALYSIS] establishes that an attacker gains an advantage over an ideal PRP of no more than:

v / 2^t + (2l * (v + q))^2 / 2^n

The goal is to limit this advantage to 2^-57, to match the target in [TLS13]. As t and n are both 128, the first term is negligible relative to the second, so that term can be removed without a significant effect on the result. This produces the relation:

v + q <= 2^24.5

Using the previously-established value of 2^23 for q and rounding, this leads to an upper limit on v of 2^23.5. That is, endpoints cannot attempt to authenticate more than 2^23.5 packets with the same set of keys without causing an attacker to gain an larger advantage than the target of 2^-57.

B.3. Limits for AEAD_AES_128_CCM_8

The TLS_AES_128_CCM_8_SHA256 cipher suite uses the AEAD_AES_128_CCM_8 function, which uses a short authentication tag (that is, t=64).

The confidentiality limits of AEAD_AES_128_CCM_8 are the same as those for AEAD_AES_128_CCM, as this does not depend on the tag length; see Appendix B.1.

The shorter tag length of 64 bits means that the simplification used in Appendix B.2 does not apply to AEAD_AES_128_CCM_8. If the goal is to preserve the same margins as other cipher suites, then the limit on forgeries is largely dictated by the first term of the advantage formula:

v <= 2^7

As this represents attempts to fail authentication, applying this limit might be feasible in some environments. However, applying this limit in an implementation intended for general use exposes connections to an inexpensive denial of service attack.

This analysis supports the view that TLS_AES_128_CCM_8_SHA256 is not suitable for general use. Specifically, TLS_AES_128_CCM_8_SHA256 cannot be used without additional measures to prevent forgery of records, or to mitigate the effect of forgeries. This might require understanding the constraints that exist in a particular deployment or application. For instance, it might be possible to set a different target for the advantage an attacker gains based on an understanding of the constraints imposed on a specific usage of DTLS.

Appendix C. History


IETF Drafts

draft-39 - Updated Figure 4 due to misalignment with Figure 3 content

draft-38 - Ban implicit connection IDs (*) - ACKs are processed as the union.

draft-37: - Fix the other place where we have ACK.

draft-36: - Some editorial changes. - Changed the content type to not conflict with existing allocations (*)

draft-35: - I-D.ietf-tls-dtls-connection-id became a normative reference - Removed duplicate reference to I-D.ietf-tls-dtls-connection-id. - Fix figure 11 to have the right numbers andno cookie in message 1. - Clarify when you can ACK. - Clarify additional data computation.

draft-33: - Key separation between TLS and DTLS. Issue #72.

draft-32: - Editorial improvements and clarifications.

draft-31: - Editorial improvements in text and figures. - Added normative reference to ChaCha20 and Poly1305.

draft-30: - Changed record format - Added text about end of early data - Changed format of the Connection ID Update message - Added Appendix A "Protocol Data Structures and Constant Values"

draft-29: - Added support for sequence number encryption - Update to new record format - Emphasize that compatibility mode isn't used.

draft-28: - Version bump to align with TLS 1.3 pre-RFC version.

draft-27: - Incorporated unified header format. - Added support for CIDs.

draft-04 - 26: - Submissions to align with TLS 1.3 draft versions

draft-03 - Only update keys after KeyUpdate is ACKed.

draft-02 - Shorten the protected record header and introduce an ultra-short version of the record header. - Reintroduce KeyUpdate, which works properly now that we have ACK. - Clarify the ACK rules.

draft-01 - Restructured the ACK to contain a list of records and also be a record rather than a handshake message.

draft-00 - First IETF Draft

Personal Drafts draft-01 - Alignment with version -19 of the TLS 1.3 specification


Appendix D. Working Group Information


The discussion list for the IETF TLS working group is located at the e-mail address Information on the group and information on how to subscribe to the list is at

Archives of the list can be found at:

Appendix E. Contributors

Many people have contributed to previous DTLS versions and they are acknowledged in prior versions of DTLS specifications or in the referenced specifications. The sequence number encryption concept is taken from the QUIC specification. We would like to thank the authors of the QUIC specification for their work. Felix Guenther and Martin Thomson contributed the analysis in Appendix B.

In addition, we would like to thank:

* David Benjamin
* Thomas Fossati
  Arm Limited
* Tobias Gondrom
* Felix Günther
  ETH Zurich
* Ilari Liusvaara
* Martin Thomson
* Christopher A. Wood
  Apple Inc.
* Yin Xinxing
* Hanno Becker
  Arm Limited

Appendix F. Acknowledgements

We would like to thank Jonathan Hammell for his review comments.

Authors' Addresses

Eric Rescorla
RTFM, Inc.
Hannes Tschofenig
Arm Limited
Nagendra Modadugu
Google, Inc.