curl/docs/CONNECTION-FILTERS.md

curl connection filters

Connection filters is a design in the internals of curl, not visible in its public API. They were added in curl v7.87.0. This document describes the concepts, its high level implementation and the motivations.

Filters

A "connection filter" is a piece of code that is responsible for handling a range of operations of curl's connections: reading, writing, waiting on external events, connecting and closing down - to name the most important ones.

The most important feat of connection filters is that they can be stacked on top of each other (or "chained" if you prefer that metaphor). In the common scenario that you want to retrieve a https: URL with curl, you need 2 basic things to send the request and get the response: a TCP connection, represented by a socket and a SSL instance en- and decrypt over that socket. You write your request to the SSL instance, which encrypts and writes that data to the socket, which then sends the bytes over the network.

With connection filters, curl's internal setup looks something like this (cf for connection filter):

Curl_easy *data         connectdata *conn        cf-ssl        cf-socket
+----------------+      +-----------------+      +-------+     +--------+
|https://curl.se/|----> | properties      |----> | keys  |---> | socket |--> OS --> network
+----------------+      +-----------------+      +-------+     +--------+

 Curl_write(data, buffer)
  --> Curl_cfilter_write(data, data->conn, buffer)
       ---> conn->filter->write(conn->filter, data, buffer)

While connection filters all do different things, they look the same from the "outside". The code in data and conn does not really know which filters are installed. conn just writes into the first filter, whatever that is.

Same is true for filters. Each filter has a pointer to the next filter. When SSL has encrypted the data, it does not write to a socket, it writes to the next filter. If that is indeed a socket, or a file, or an HTTP/2 connection is of no concern to the SSL filter.

This allows stacking, as in:

Direct:
  http://localhost/      conn -> cf-socket
  https://curl.se/       conn -> cf-ssl -> cf-socket
Via http proxy tunnel:
  http://localhost/      conn -> cf-http-proxy -> cf-socket
  https://curl.se/       conn -> cf-ssl -> cf-http-proxy -> cf-socket
Via https proxy tunnel:
  http://localhost/      conn -> cf-http-proxy -> cf-ssl -> cf-socket
  https://curl.se/       conn -> cf-ssl -> cf-http-proxy -> cf-ssl -> cf-socket
Via http proxy tunnel via SOCKS proxy:
  http://localhost/      conn -> cf-http-proxy -> cf-socks -> cf-socket

Connecting/Closing

Before Curl_easy can send the request, the connection needs to be established. This means that all connection filters have done, whatever they need to do: waiting for the socket to be connected, doing the TLS handshake, performing the HTTP tunnel request, etc. This has to be done in reverse order: the last filter has to do its connect first, then the one above can start, etc.

Each filter does in principle the following:

static CURLcode
myfilter_cf_connect(struct Curl_cfilter *cf,
                    struct Curl_easy *data,
                    bool *done)
{
  CURLcode result;

  if(cf->connected) {            /* we and all below are done */
    *done = TRUE;
    return CURLE_OK;
  }
                                 /* Let the filters below connect */
  result = cf->next->cft->connect(cf->next, data, blocking, done);
  if(result || !*done)
    return result;               /* below errored/not finished yet */

  /* MYFILTER CONNECT THINGS */  /* below connected, do out thing */
  *done = cf->connected = TRUE;  /* done, remember, return */
  return CURLE_OK;
}

Closing a connection then works similar. The conn tells the first filter to close. Contrary to connecting, the filter does its own things first, before telling the next filter to close.

Efficiency

There are two things curl is concerned about: efficient memory use and fast transfers.

The memory footprint of a filter is relatively small:

struct Curl_cfilter {
  const struct Curl_cftype *cft; /* the type providing implementation */
  struct Curl_cfilter *next;     /* next filter in chain */
  void *ctx;                     /* filter type specific settings */
  struct connectdata *conn;      /* the connection this filter belongs to */
  int sockindex;                 /* TODO: like to get rid off this */
  BIT(connected);                /* != 0 iff this filter is connected */
};

The filter type cft is a singleton, one static struct for each type of filter. The ctx is where a filter holds its specific data. That varies by filter type. An http-proxy filter keeps the ongoing state of the CONNECT here, free it after its has been established. The SSL filter keeps the SSL* (if OpenSSL is used) here until the connection is closed. So, this varies.

conn is a reference to the connection this filter belongs to, so nothing extra besides the pointer itself.

Several things, that before were kept in struct connectdata, now goes into the filter->ctx when needed. So, the memory footprint for connections that do not use an http proxy, or socks, or https is lower.

As to transfer efficiency, writing and reading through a filter comes at near zero cost if the filter does not transform the data. An http proxy or socks filter, once it is connected, just passes the calls through. Those filters implementations look like this:

ssize_t  Curl_cf_def_send(struct Curl_cfilter *cf, struct Curl_easy *data,
                          const void *buf, size_t len, CURLcode *err)
{
  return cf->next->cft->do_send(cf->next, data, buf, len, err);
}

The recv implementation is equivalent.

Filter Types

The currently existing filter types (curl 8.5.0) are:

• TCP, UDP, UNIX: filters that operate on a socket, providing raw I/O.
• SOCKET-ACCEPT: special TCP socket that has a socket that has been accept()ed in a listen()
• SSL: filter that applies TLS en-/decryption and handshake. Manages the underlying TLS backend implementation.
• HTTP-PROXY, H1-PROXY, H2-PROXY: the first manages the connection to an HTTP proxy server and uses the other depending on which ALPN protocol has been negotiated.
• SOCKS-PROXY: filter for the various SOCKS proxy protocol variations
• HAPROXY: filter for the protocol of the same name, providing client IP information to a server.
• HTTP/2: filter for handling multiplexed transfers over an HTTP/2 connection
• HTTP/3: filter for handling multiplexed transfers over an HTTP/3+QUIC connection
• HAPPY-EYEBALLS: meta filter that implements IPv4/IPv6 "happy eyeballing". It creates up to 2 sub-filters that race each other for a connection.
• SETUP: meta filter that manages the creation of sub-filter chains for a specific transport (e.g. TCP or QUIC).
• HTTPS-CONNECT: meta filter that races a TCP+TLS and a QUIC connection against each other to determine if HTTP/1.1, HTTP/2 or HTTP/3 shall be used for a transfer.

Meta filters are combining other filters for a specific purpose, mostly during connection establishment. Other filters like TCP, UDP and UNIX are only to be found at the end of filter chains. SSL filters provide encryption, of course. Protocol filters change the bytes sent and received.

Filter Flags

Filter types carry flags that inform what they do. These are (for now):

• CF_TYPE_IP_CONNECT: this filter type talks directly to a server. This does not have to be the server the transfer wants to talk to. For example when a proxy server is used.
• CF_TYPE_SSL: this filter type provides encryption.
• CF_TYPE_MULTIPLEX: this filter type can manage multiple transfers in parallel.

Filter types can combine these flags. For example, the HTTP/3 filter types have CF_TYPE_IP_CONNECT, CF_TYPE_SSL and CF_TYPE_MULTIPLEX set.

Flags are useful to extrapolate properties of a connection. To check if a connection is encrypted, libcurl inspect the filter chain in place, top down, for CF_TYPE_SSL. If it finds CF_TYPE_IP_CONNECT before any CF_TYPE_SSL, the connection is not encrypted.

For example, conn1 is for a http: request using a tunnel through an HTTP/2 https: proxy. conn2 is a https: HTTP/2 connection to the same proxy. conn3 uses HTTP/3 without proxy. The filter chains would look like this (simplified):

conn1 --> `HTTP-PROXY` --> `H2-PROXY` --> `SSL` --> `TCP`
flags:                     `IP_CONNECT`   `SSL`     `IP_CONNECT`

conn2 --> `HTTP/2` --> `SSL` --> `HTTP-PROXY` --> `H2-PROXY` --> `SSL` --> `TCP`
flags:                 `SSL`                      `IP_CONNECT`   `SSL`     `IP_CONNECT`

conn3 --> `HTTP/3`
flags:    `SSL|IP_CONNECT`

Inspecting the filter chains, conn1 is seen as unencrypted, since it contains an IP_CONNECT filter before any SSL. conn2 is clearly encrypted as an SSL flagged filter is seen first. conn3 is also encrypted as the SSL flag is checked before the presence of IP_CONNECT.

Similar checks can determine if a connection is multiplexed or not.

Filter Tracing

Filters may make use of special trace macros like CURL_TRC_CF(data, cf, msg, ...). With data being the transfer and cf being the filter instance. These traces are normally not active and their execution is guarded so that they are cheap to ignore.

Users of curl may activate them by adding the name of the filter type to the --trace-config argument. For example, in order to get more detailed tracing of an HTTP/2 request, invoke curl with:

> curl -v --trace-config ids,time,http/2  https://curl.se

Which gives you trace output with time information, transfer+connection ids and details from the HTTP/2 filter. Filter type names in the trace config are case insensitive. You may use all to enable tracing for all filter types. When using libcurl you may call curl_global_trace(config_string) at the start of your application to enable filter details.

Meta Filters

Meta filters is a catch-all name for filter types that do not change the transfer data in any way but provide other important services to curl. In general, it is possible to do all sorts of silly things with them. One of the commonly used, important things is "eyeballing".

The HAPPY-EYEBALLS filter is involved in the connect phase. Its job is to try the various IPv4 and IPv6 addresses that are known for a server. If only one address family is known (or configured), it tries the addresses one after the other with timeouts calculated from the amount of addresses and the overall connect timeout.

When more than one address family is to be tried, it splits the address list into IPv4 and IPv6 and makes parallel attempts. The connection filter chain looks like this:

* create connection for http://curl.se
conn[curl.se] --> SETUP[TCP] --> HAPPY-EYEBALLS --> NULL
* start connect
conn[curl.se] --> SETUP[TCP] --> HAPPY-EYEBALLS --> NULL
                                 - ballerv4 --> TCP[151.101.1.91]:443
                                 - ballerv6 --> TCP[2a04:4e42:c00::347]:443
* v6 answers, connected
conn[curl.se] --> SETUP[TCP] --> HAPPY-EYEBALLS --> TCP[2a04:4e42:c00::347]:443
* transfer

The modular design of connection filters and that we can plug them into each other is used to control the parallel attempts. When a TCP filter does not connect (in time), it is torn down and another one is created for the next address. This keeps the TCP filter simple.

The HAPPY-EYEBALLS on the other hand stays focused on its side of the problem. We can use it also to make other type of connection by just giving it another filter type to try to have happy eyeballing for QUIC:

* create connection for --http3-only https://curl.se
conn[curl.se] --> SETUP[QUIC] --> HAPPY-EYEBALLS --> NULL
* start connect
conn[curl.se] --> SETUP[QUIC] --> HAPPY-EYEBALLS --> NULL
                                  - ballerv4 --> HTTP/3[151.101.1.91]:443
                                  - ballerv6 --> HTTP/3[2a04:4e42:c00::347]:443
* v6 answers, connected
conn[curl.se] --> SETUP[QUIC] --> HAPPY-EYEBALLS --> HTTP/3[2a04:4e42:c00::347]:443
* transfer

When we plug these two variants together, we get the HTTPS-CONNECT filter type that is used for --http3 when both HTTP/3 and HTTP/2 or HTTP/1.1 shall be attempted:

* create connection for --http3 https://curl.se
conn[curl.se] --> HTTPS-CONNECT --> NULL
* start connect
conn[curl.se] --> HTTPS-CONNECT --> NULL
                  - SETUP[QUIC] --> HAPPY-EYEBALLS --> NULL
                                    - ballerv4 --> HTTP/3[151.101.1.91]:443
                                    - ballerv6 --> HTTP/3[2a04:4e42:c00::347]:443
                  - SETUP[TCP]  --> HAPPY-EYEBALLS --> NULL
                                    - ballerv4 --> TCP[151.101.1.91]:443
                                    - ballerv6 --> TCP[2a04:4e42:c00::347]:443
* v4 QUIC answers, connected
conn[curl.se] --> HTTPS-CONNECT --> SETUP[QUIC] --> HAPPY-EYEBALLS --> HTTP/3[151.101.1.91]:443
* transfer