mnot@mnot.net https://www.mnot.net/

General Internet-Draft HTTP allows requests to be automatically retried under certain circumstances. This draft explores how this is implemented, requirements for similar functionality from other parts of the stack, and potential future improvements. This draft is not intended to be published as an RFC. The issues list for this draft can be found at https://github.com/mnot/I-D/labels/httpbis-retry. The most recent (often, unpublished) draft is at https://mnot.github.io/I-D/httpbis-retry/. Recent changes are listed at https://github.com/mnot/I-D/commits/gh-pages/httpbis-retry.

One of the benefits of HTTP’s well-defined method semantics is that they allow failed requests to be retried, under certain circumstances. However, interest in extending, redefining or just clarifying HTTP’s retry semantics is increasing, for a number of reasons: Since HTTP/1.1’s requirements were written, there has been a substantial amount of experience deploying and using HTTP, leading implementations to refine their behaviour, arguably diverging from the specification. Likewise, changes such as HTTP/2 might change the underlying assumptions that these requirements were based upon. Emerging lower-layer developments such as TCP Fast Open , TLS/1.3 and QUIC introduce the possibility of replayed requests in the beginning of a connection, thanks to Zero Round Trip (0RT) modes. In some ways, these are similar to retries – but not completely. Applications sometimes want requests to be retried by infrastructure, but can’t easily express them in a non-idempotent request (such as GET). This draft gives some background in , discusses aspects of these issues in , suggesting possible areas of work in , and cataloguing current implementation behaviours in .

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 .

In HTTP, there are three similar but separate phenomena that deserve consideration for the purposes of this document: User Retries happen when a user initiates an action that results in a duplicate request being emitted. For example, a user retry might occur when a “reload” button is pressed, a URL is typed in again, “return” is pressed in the URL bar again, or a navigation link or form button is pressed twice while still on screen. Automatic Retries happen when an HTTP client implementation resends a previous request without user intervention or initiation. This might happen when a GET request fails to return a complete response, or when a connection drops before the request is sent. Note that automatic retries can (and are) performed both by user agents and intermediary clients. Replays happen when the packet(s) containing a request are re-sent on the network, either automatically as part of transport protocol operation, or by an attacker. The closest upstream HTTP client might not have any indication that a replay has occurred. Note that retries initiated by code shipped to the client by the server (e.g., in JavaScript) occupy a grey area here; however, because they are not initiated by the generic HTTP client implementation itself, we will consider them user retries for the time being. Also, this document doesn’t include TCP-layer loss recovery (i.e., retransmission).

, Section 6.3.1 allows HTTP requests to be retried in certain circumstances: When an inbound connection is closed prematurely, a client MAY open a new connection and automatically retransmit an aborted sequence of requests if all of those requests have idempotent methods (Section 4.2.2 of ). A proxy MUST NOT automatically retry non-idempotent requests. A user agent MUST NOT automatically retry a request with a non-idempotent method unless it has some means to know that the request semantics are actually idempotent, regardless of the method, or some means to detect that the original request was never applied. For example, a user agent that knows (through design or configuration) that a POST request to a given resource is safe can repeat that request automatically. Likewise, a user agent designed specifically to operate on a version control repository might be able to recover from partial failure conditions by checking the target resource revision(s) after a failed connection, reverting or fixing any changes that were partially applied, and then automatically retrying the requests that failed. A client SHOULD NOT automatically retry a failed automatic retry. Note that the complete list of idempotent methods is maintained in the IANA HTTP Method Registry.

, Section 6.3.1 addresses HTTP Request Replay with TCP Fast Open: While TFO is motivated by Web applications, the browser should not use TFO to send requests in SYNs if those requests cannot tolerate replays. One example is POST requests without application-layer transaction protection (e.g., a unique identifier in the request header). On the other hand, TFO is particularly useful for GET requests. GET request replay could happen across striped TCP connections: after a server receives an HTTP request but before the ACKs of the requests reach the browser, the browser may time out and retry the same request on another (possibly new) TCP connection. This differs from a TFO replay only in that the replay is initiated by the browser, not by the TCP stack. The same specification addresses HTTP over TLS in Section 6.3.2: For Transport Layer Security (TLS) over TCP, it is safe and useful to include a TLS client_hello in the SYN packet to save one RTT in the TLS handshake. There is no concern about violating idempotency. In particular, it can be used alone with the speculative connection above.

, Section 2.3 explains the properties of Zero-RTT Data in TLS 1.3: IMPORTANT NOTE: The security properties for 0-RTT data (regardless of the cipher suite) are weaker than those for other kinds of TLS data. Specifically: This data is not forward secret, because it is encrypted solely with the PSK. There are no guarantees of non-replay between connections. Unless the server takes special measures outside those provided by TLS, the server has no guarantee that the same 0-RTT data was not transmitted on multiple 0-RTT connections (See Section 4.2.6.2 for more details). This is especially relevant if the data is authenticated either with TLS client authentication or inside the application layer protocol. However, 0-RTT data cannot be duplicated within a connection (i.e., the server will not process the same data twice for the same connection) and an attacker will not be able to make 0-RTT data appear to be 1-RTT data (because it is protected with different keys.) Section 4.2.6 defines a mechanism to limit the exposure to replay.

The QUIC specifications don’t say anything about the replay risk of 0RTT.

In practice, it has been observed (see ) that some client (both user agent and intermediary) implementations do automatically retry requests. However, they do not do so consistently, and arguably not in the spirit of the specification, unless this vague catch-all: some means to detect that the original request was never applied is interpreted very broadly. On the server side, it has been widely observed that content on the Web doesn’t always honour HTTP idemotency semantics, with many GET requests incurring side effects, and with some sites even requiring browsers to retry POST requests in order to properly interoperate. (TODO: refs / details from Patrick and Artur). Despite this situation, the Web seems to work reasonably well to date (with notable exceptions). The status quo, therefore, is that no Web application can read HTTP’s retry requirements as a guarantee that any given request won’t be retried, even for methods that are not idempotent. As a result, applications that care about avoiding duplicate requests need to build a way to detect not only user retries but also automatic retries into the application “above” HTTP itself.

TCP Fast Open , TLS/1.3 and QUIC all have mechanisms to carry application data on the first packet sent by a client, to avoid the latency of connection setup. The request(s) in this first packet might be replayed, either because the first packet is lost and retransmitted by the transport protocol in use, or because an attacker observes the packet and sends a duplicate at some point in the future. At first glance, it seems as if the idempotency semantics of HTTP request methods could be used to determine what requests are suitable for inclusion in the first packet of various 0RT mechanisms being discussed (as suggested by TCP Fast Open). For example, we could disallow POST (and other non-idempotent methods) in 0RT data. Upon reflection, though, the observations above lead us to believe that since any request might be retried (automatically or by users), applications will need to have a means of detecting duplicate requests, thereby preventing side effects from replays as well as retries. Thus, any HTTP request can be included in the first packet of a 0RT, despite the risk of replay. Two types of attack specific to replayed HTTP requests need to be taken into account, however: A replay is a potential Denial of Service vector. An attacker that can replay a request many times can probe for weaknesses in retry protections, and can bring a server that needs to do any substantial processing down. An attacker might use a replayed request to leak information about the response over time. If they can observe the encrypted payload on the wire, they can infer the size of the response (e.g., it might get bigger if the user’s bank account has more in it). The first attack cannot be mitigated by HTTP; the 0RT mechanism itself needs some transport-layer means of scoping the usability of the first packet on a connection so that it cannot be reused broadly. For example, this might be by time, network location. The second attack is more difficult to mitigate; scoping the usability of the first packet helps, but does not completely prevent the attack. If the replayed request is state-changing, the application’s retry detection should kick in and prevent information leakage (since the response will likely contain an error, instead of the desired information). If it is not (e.g., a GET), the information being targeted is vulnerable as long as both the first packet and the credentials in the request (if any) are valid.

The currently language in about retries is vague about the conditions under which a request can be retried, leading to significant variance in implementation behaviour. For example, it’s been observed that many automated clients fail under circumstances when browsers succeed, because they do not retry in the same way. As a result, more carefully specifying the conditions under which a request can be retried would be helpful. Such work would need to take into account varying conditions, such as: Connection closes TCP RST Connection timeouts Whether or not any part of the response has been received Whether or not it is the first request on the connection Variance due to use of HTTP/2, TLS/1.3 and TCP Fast Open. Furthermore, readers might mistake the language in RFC7230 as guaranteeing that some requests (e.g., POST) are never automatically retried; this should be clarified.

A number of mechanisms have been mooted at various times, e.g.: Adding a header to automatically retried requests, to aid de-duplication by servers Defining a request header to by added by intermediaries when they have received a request in a way that could have been replayed Defining a status code to allow servers to indicate that the request needs to be sent in a way that can’t be replayed

If the observations above hold, we should disabuse any notion that HTTP method idempotency is a useful way to avoid problems with replay attacks. Instead, we should encourage development of mechanisms to mitigate the aspects of replay that are different than retries (e.g., potential for DOS attacks).

Yep.

Thanks to Brad Fitzpatrick, Leif Hedstrom, Subodh Iyengar, Amos Jeffries, Patrick McManus, Matt Menke, Miroslav Ponec and Martin Thomson for their input and feedback. Thanks also to the participants in the 2016 HTTP Workshop for their lively discussion of this topic.

This document describes an experimental TCP mechanism called TCP Fast Open (TFO). TFO allows data to be carried in the SYN and SYN-ACK packets and consumed by the receiving end during the initial connection handshake, and saves up to one full round-trip time (RTT) compared to the standard TCP, which requires a three-way handshake (3WHS) to complete before data can be exchanged. However, TFO deviates from the standard TCP semantics, since the data in the SYN could be replayed to an application in some rare circumstances. Applications should not use TFO unless they can tolerate this issue, as detailed in the Applicability section. In many standards track documents several words are used to signify the requirements in the specification. These words are often capitalized. This document defines these words as they should be interpreted in IETF documents. This document specifies an Internet Best Current Practices for the Internet Community, and requests discussion and suggestions for improvements. The Hypertext Transfer Protocol (HTTP) is a stateless application-level protocol for distributed, collaborative, hypertext information systems. This document provides an overview of HTTP architecture and its associated terminology, defines the "http" and "https" Uniform Resource Identifier (URI) schemes, defines the HTTP/1.1 message syntax and parsing requirements, and describes related security concerns for implementations. The Hypertext Transfer Protocol (HTTP) is a stateless \%application- level protocol for distributed, collaborative, hypertext information systems. This document defines the semantics of HTTP/1.1 messages, as expressed by request methods, request header fields, response status codes, and response header fields, along with the payload of messages (metadata and body content) and mechanisms for content negotiation. This specification describes an optimized expression of the semantics of the Hypertext Transfer Protocol (HTTP), referred to as HTTP version 2 (HTTP/2). HTTP/2 enables a more efficient use of network resources and a reduced perception of latency by introducing header field compression and allowing multiple concurrent exchanges on the same connection. It also introduces unsolicited push of representations from servers to clients.This specification is an alternative to, but does not obsolete, the HTTP/1.1 message syntax. HTTP's existing semantics remain unchanged. This document specifies version 1.3 of the Transport Layer Security (TLS) protocol. TLS allows client/server applications to communicate over the Internet in a way that is designed to prevent eavesdropping, tampering, and message forgery. QUIC (Quick UDP Internet Connection) is a new multiplexed and secure transport atop UDP, designed from the ground up and optimized for HTTP/2 semantics. While built with HTTP/2 as the primary application protocol, QUIC builds on decades of transport and security experience, and implements mechanisms that make it attractive as a modern general-purpose transport. QUIC provides multiplexing and flow control equivalent to HTTP/2, security equivalent to TLS, and connection semantics, reliability, and congestion control equivalent to TCP. This draft documents the early deployment of the QUIC protocol prior to standardization.

In implementations, clients have been observed to retry requests in a number of circumstances. Note: This section is intended to inform the discussion, not to be published as a standard. If you have relevant information about these or other implementations (open or closed), please get in touch.

Squid is a caching proxy server that retries requests that it considers safe or idempotent, as long as there is not a request body:

body_pipe != NULL) return false; // RFC2616 9.1 Safe and Idempotent Methods return (request->method.isHttpSafe() || request->method.isIdempotent()); } ]]>

(source) Currently, it considers GET, HEAD, OPTIONS, REPORT, PROPFIND, SEARCH and PRI to be safe, and GET, HEAD, PUT, DELETE, OPTIONS, TRACE, PROPFIND, PROPPATCH, MKCOL, COPY, MOVE, UNLOCK, and PRI to be idempotent.

Apache Traffic Server, a caching proxy server, ties retry-ability to whether the request required a “tunnel” – i.e., forward to the next server. This is indicated by request_body_start, which is set when a POST tunnel is used.

hdr_info.request_body_start == true) { return false; } if (s->state_machine->plugin_tunnel_type != HTTP_NO_PLUGIN_TUNNEL) { // API can override if (s->state_machine->plugin_tunnel_type == HTTP_PLUGIN_AS_SERVER && s->api_info.retry_intercept_failures == true) { // This used to be an == comparison, which made no sense. Changed // to be an assignment, hoping the state is correct. s->state_machine->plugin_tunnel_type = HTTP_NO_PLUGIN_TUNNEL; } else { return false; } } return true; } ]]>

(source) When connected to an origin server, Traffic Server attempts to retry under a number of failure conditions:

current.state) { case CONNECTION_ALIVE: DebugTxn("http_trans", "[hrfs] connection alive"); SET_VIA_STRING(VIA_DETAIL_SERVER_CONNECT, VIA_DETAIL_SERVER_SUCCESS); s->current.server->clear_connect_fail(); handle_forward_server_connection_open(s); break; [...] case OPEN_RAW_ERROR: /* fall through */ case CONNECTION_ERROR: /* fall through */ case STATE_UNDEFINED: /* fall through */ case INACTIVE_TIMEOUT: // Set to generic I/O error if not already set specifically. if (!s->current.server->had_connect_fail()) s->current.server->set_connect_fail(EIO); if (is_server_negative_cached(s)) { max_connect_retries = s->txn_conf->connect_attempts_max_retries_dead_server; } else { // server not yet negative cached - use default number of retries max_connect_retries = s->txn_conf->connect_attempts_max_retries; } if (s->pCongestionEntry != NULL) max_connect_retries = s->pCongestionEntry->connect_retries(); if (is_request_retryable(s) && s->current.attempts < max_connect_retries) { ]]>

(source)

Firefox is a Web browser that retries under the following conditions:

IsSafeMethod()) || !reallySentData || connReused)) { // if restarting fails, then we must proceed to close the pipe, // which will notify the channel that the transaction failed. ]]>

(source) … and it considers GET, HEAD, OPTIONS, TRACE, PROPFIND, REPORT, and SEARCH to be safe:

(source) Note that connReused is tested; if a connection has been used before, Firefox will retry any request, safe or not. A recent change attempted to remove this behaviour, but it caused compatibility problems, and is being backed out.

Chromium is a Web browser that appears to retry any request when a connection is broken, as long as it’s successfully used the connection before, and hasn’t received any response headers yet:

IsConnectionReused(); bool has_received_headers = GetResponseHeaders() != NULL; // NOTE: we resend a request only if we reused a keep-alive connection. // This automatically prevents an infinite resend loop because we'll run // out of the cached keep-alive connections eventually. if (connection_is_proven && !has_received_headers) return true; return false; } ]]>

(source)