1<!-- 2Copyright (C) Daniel Stenberg, <daniel@haxx.se>, et al. 3 4SPDX-License-Identifier: curl 5--> 6 7# curl connection filters 8 9Connection filters is a design in the internals of curl, not visible in its 10public API. They were added in curl v7.87.0. This document describes the 11concepts, its high level implementation and the motivations. 12 13## Filters 14 15A "connection filter" is a piece of code that is responsible for handling a 16range of operations of curl's connections: reading, writing, waiting on 17external events, connecting and closing down - to name the most important 18ones. 19 20The most important feat of connection filters is that they can be stacked on 21top of each other (or "chained" if you prefer that metaphor). In the common 22scenario that you want to retrieve a `https:` URL with curl, you need 2 basic 23things to send the request and get the response: a TCP connection, represented 24by a `socket` and a SSL instance en- and decrypt over that socket. You write 25your request to the SSL instance, which encrypts and writes that data to the 26socket, which then sends the bytes over the network. 27 28With connection filters, curl's internal setup looks something like this (cf 29for connection filter): 30 31``` 32Curl_easy *data connectdata *conn cf-ssl cf-socket 33+----------------+ +-----------------+ +-------+ +--------+ 34|https://curl.se/|----> | properties |----> | keys |---> | socket |--> OS --> network 35+----------------+ +-----------------+ +-------+ +--------+ 36 37 Curl_write(data, buffer) 38 --> Curl_cfilter_write(data, data->conn, buffer) 39 ---> conn->filter->write(conn->filter, data, buffer) 40``` 41 42While connection filters all do different things, they look the same from the 43"outside". The code in `data` and `conn` does not really know **which** 44filters are installed. `conn` just writes into the first filter, whatever that 45is. 46 47Same is true for filters. Each filter has a pointer to the `next` filter. When 48SSL has encrypted the data, it does not write to a socket, it writes to the 49next filter. If that is indeed a socket, or a file, or an HTTP/2 connection is 50of no concern to the SSL filter. 51 52This allows stacking, as in: 53 54``` 55Direct: 56 http://localhost/ conn -> cf-socket 57 https://curl.se/ conn -> cf-ssl -> cf-socket 58Via http proxy tunnel: 59 http://localhost/ conn -> cf-http-proxy -> cf-socket 60 https://curl.se/ conn -> cf-ssl -> cf-http-proxy -> cf-socket 61Via https proxy tunnel: 62 http://localhost/ conn -> cf-http-proxy -> cf-ssl -> cf-socket 63 https://curl.se/ conn -> cf-ssl -> cf-http-proxy -> cf-ssl -> cf-socket 64Via http proxy tunnel via SOCKS proxy: 65 http://localhost/ conn -> cf-http-proxy -> cf-socks -> cf-socket 66``` 67 68### Connecting/Closing 69 70Before `Curl_easy` can send the request, the connection needs to be 71established. This means that all connection filters have done, whatever they 72need to do: waiting for the socket to be connected, doing the TLS handshake, 73performing the HTTP tunnel request, etc. This has to be done in reverse order: 74the last filter has to do its connect first, then the one above can start, 75etc. 76 77Each filter does in principle the following: 78 79``` 80static CURLcode 81myfilter_cf_connect(struct Curl_cfilter *cf, 82 struct Curl_easy *data, 83 bool *done) 84{ 85 CURLcode result; 86 87 if(cf->connected) { /* we and all below are done */ 88 *done = TRUE; 89 return CURLE_OK; 90 } 91 /* Let the filters below connect */ 92 result = cf->next->cft->connect(cf->next, data, blocking, done); 93 if(result || !*done) 94 return result; /* below errored/not finished yet */ 95 96 /* MYFILTER CONNECT THINGS */ /* below connected, do out thing */ 97 *done = cf->connected = TRUE; /* done, remember, return */ 98 return CURLE_OK; 99} 100``` 101 102Closing a connection then works similar. The `conn` tells the first filter to 103close. Contrary to connecting, the filter does its own things first, before 104telling the next filter to close. 105 106### Efficiency 107 108There are two things curl is concerned about: efficient memory use and fast 109transfers. 110 111The memory footprint of a filter is relatively small: 112 113``` 114struct Curl_cfilter { 115 const struct Curl_cftype *cft; /* the type providing implementation */ 116 struct Curl_cfilter *next; /* next filter in chain */ 117 void *ctx; /* filter type specific settings */ 118 struct connectdata *conn; /* the connection this filter belongs to */ 119 int sockindex; /* TODO: like to get rid off this */ 120 BIT(connected); /* != 0 iff this filter is connected */ 121}; 122``` 123 124The filter type `cft` is a singleton, one static struct for each type of 125filter. The `ctx` is where a filter holds its specific data. That varies by 126filter type. An http-proxy filter keeps the ongoing state of the CONNECT here, 127free it after its has been established. The SSL filter keeps the `SSL*` (if 128OpenSSL is used) here until the connection is closed. So, this varies. 129 130`conn` is a reference to the connection this filter belongs to, so nothing 131extra besides the pointer itself. 132 133Several things, that before were kept in `struct connectdata`, now goes into 134the `filter->ctx` *when needed*. So, the memory footprint for connections that 135do *not* use an http proxy, or socks, or https is lower. 136 137As to transfer efficiency, writing and reading through a filter comes at near 138zero cost *if the filter does not transform the data*. An http proxy or socks 139filter, once it is connected, just passes the calls through. Those filters 140implementations look like this: 141 142``` 143ssize_t Curl_cf_def_send(struct Curl_cfilter *cf, struct Curl_easy *data, 144 const void *buf, size_t len, CURLcode *err) 145{ 146 return cf->next->cft->do_send(cf->next, data, buf, len, err); 147} 148``` 149The `recv` implementation is equivalent. 150 151## Filter Types 152 153The currently existing filter types (curl 8.5.0) are: 154 155* `TCP`, `UDP`, `UNIX`: filters that operate on a socket, providing raw I/O. 156* `SOCKET-ACCEPT`: special TCP socket that has a socket that has been 157 `accept()`ed in a `listen()` 158* `SSL`: filter that applies TLS en-/decryption and handshake. Manages the 159 underlying TLS backend implementation. 160* `HTTP-PROXY`, `H1-PROXY`, `H2-PROXY`: the first manages the connection to an 161 HTTP proxy server and uses the other depending on which ALPN protocol has 162 been negotiated. 163* `SOCKS-PROXY`: filter for the various SOCKS proxy protocol variations 164* `HAPROXY`: filter for the protocol of the same name, providing client IP 165 information to a server. 166* `HTTP/2`: filter for handling multiplexed transfers over an HTTP/2 167 connection 168* `HTTP/3`: filter for handling multiplexed transfers over an HTTP/3+QUIC 169 connection 170* `HAPPY-EYEBALLS`: meta filter that implements IPv4/IPv6 "happy eyeballing". 171 It creates up to 2 sub-filters that race each other for a connection. 172* `SETUP`: meta filter that manages the creation of sub-filter chains for a 173 specific transport (e.g. TCP or QUIC). 174* `HTTPS-CONNECT`: meta filter that races a TCP+TLS and a QUIC connection 175 against each other to determine if HTTP/1.1, HTTP/2 or HTTP/3 shall be used 176 for a transfer. 177 178Meta filters are combining other filters for a specific purpose, mostly during 179connection establishment. Other filters like `TCP`, `UDP` and `UNIX` are only 180to be found at the end of filter chains. SSL filters provide encryption, of 181course. Protocol filters change the bytes sent and received. 182 183## Filter Flags 184 185Filter types carry flags that inform what they do. These are (for now): 186 187* `CF_TYPE_IP_CONNECT`: this filter type talks directly to a server. This does 188 not have to be the server the transfer wants to talk to. For example when a 189 proxy server is used. 190* `CF_TYPE_SSL`: this filter type provides encryption. 191* `CF_TYPE_MULTIPLEX`: this filter type can manage multiple transfers in parallel. 192 193Filter types can combine these flags. For example, the HTTP/3 filter types 194have `CF_TYPE_IP_CONNECT`, `CF_TYPE_SSL` and `CF_TYPE_MULTIPLEX` set. 195 196Flags are useful to extrapolate properties of a connection. To check if a 197connection is encrypted, libcurl inspect the filter chain in place, top down, 198for `CF_TYPE_SSL`. If it finds `CF_TYPE_IP_CONNECT` before any `CF_TYPE_SSL`, 199the connection is not encrypted. 200 201For example, `conn1` is for a `http:` request using a tunnel through an HTTP/2 202`https:` proxy. `conn2` is a `https:` HTTP/2 connection to the same proxy. 203`conn3` uses HTTP/3 without proxy. The filter chains would look like this 204(simplified): 205 206``` 207conn1 --> `HTTP-PROXY` --> `H2-PROXY` --> `SSL` --> `TCP` 208flags: `IP_CONNECT` `SSL` `IP_CONNECT` 209 210conn2 --> `HTTP/2` --> `SSL` --> `HTTP-PROXY` --> `H2-PROXY` --> `SSL` --> `TCP` 211flags: `SSL` `IP_CONNECT` `SSL` `IP_CONNECT` 212 213conn3 --> `HTTP/3` 214flags: `SSL|IP_CONNECT` 215``` 216 217Inspecting the filter chains, `conn1` is seen as unencrypted, since it 218contains an `IP_CONNECT` filter before any `SSL`. `conn2` is clearly encrypted 219as an `SSL` flagged filter is seen first. `conn3` is also encrypted as the 220`SSL` flag is checked before the presence of `IP_CONNECT`. 221 222Similar checks can determine if a connection is multiplexed or not. 223 224## Filter Tracing 225 226Filters may make use of special trace macros like `CURL_TRC_CF(data, cf, msg, 227...)`. With `data` being the transfer and `cf` being the filter instance. 228These traces are normally not active and their execution is guarded so that 229they are cheap to ignore. 230 231Users of `curl` may activate them by adding the name of the filter type to the 232`--trace-config` argument. For example, in order to get more detailed tracing 233of an HTTP/2 request, invoke curl with: 234 235``` 236> curl -v --trace-config ids,time,http/2 https://curl.se 237``` 238 239Which gives you trace output with time information, transfer+connection ids 240and details from the `HTTP/2` filter. Filter type names in the trace config 241are case insensitive. You may use `all` to enable tracing for all filter 242types. When using `libcurl` you may call `curl_global_trace(config_string)` at 243the start of your application to enable filter details. 244 245## Meta Filters 246 247Meta filters is a catch-all name for filter types that do not change the 248transfer data in any way but provide other important services to curl. In 249general, it is possible to do all sorts of silly things with them. One of the 250commonly used, important things is "eyeballing". 251 252The `HAPPY-EYEBALLS` filter is involved in the connect phase. Its job is to 253try the various IPv4 and IPv6 addresses that are known for a server. If only 254one address family is known (or configured), it tries the addresses one after 255the other with timeouts calculated from the amount of addresses and the 256overall connect timeout. 257 258When more than one address family is to be tried, it splits the address list 259into IPv4 and IPv6 and makes parallel attempts. The connection filter chain 260looks like this: 261 262``` 263* create connection for http://curl.se 264conn[curl.se] --> SETUP[TCP] --> HAPPY-EYEBALLS --> NULL 265* start connect 266conn[curl.se] --> SETUP[TCP] --> HAPPY-EYEBALLS --> NULL 267 - ballerv4 --> TCP[151.101.1.91]:443 268 - ballerv6 --> TCP[2a04:4e42:c00::347]:443 269* v6 answers, connected 270conn[curl.se] --> SETUP[TCP] --> HAPPY-EYEBALLS --> TCP[2a04:4e42:c00::347]:443 271* transfer 272``` 273 274The 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. 275 276The `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: 277 278``` 279* create connection for --http3-only https://curl.se 280conn[curl.se] --> SETUP[QUIC] --> HAPPY-EYEBALLS --> NULL 281* start connect 282conn[curl.se] --> SETUP[QUIC] --> HAPPY-EYEBALLS --> NULL 283 - ballerv4 --> HTTP/3[151.101.1.91]:443 284 - ballerv6 --> HTTP/3[2a04:4e42:c00::347]:443 285* v6 answers, connected 286conn[curl.se] --> SETUP[QUIC] --> HAPPY-EYEBALLS --> HTTP/3[2a04:4e42:c00::347]:443 287* transfer 288``` 289 290When we plug these two variants together, we get the `HTTPS-CONNECT` filter 291type that is used for `--http3` when **both** HTTP/3 and HTTP/2 or HTTP/1.1 292shall be attempted: 293 294``` 295* create connection for --http3 https://curl.se 296conn[curl.se] --> HTTPS-CONNECT --> NULL 297* start connect 298conn[curl.se] --> HTTPS-CONNECT --> NULL 299 - SETUP[QUIC] --> HAPPY-EYEBALLS --> NULL 300 - ballerv4 --> HTTP/3[151.101.1.91]:443 301 - ballerv6 --> HTTP/3[2a04:4e42:c00::347]:443 302 - SETUP[TCP] --> HAPPY-EYEBALLS --> NULL 303 - ballerv4 --> TCP[151.101.1.91]:443 304 - ballerv6 --> TCP[2a04:4e42:c00::347]:443 305* v4 QUIC answers, connected 306conn[curl.se] --> HTTPS-CONNECT --> SETUP[QUIC] --> HAPPY-EYEBALLS --> HTTP/3[151.101.1.91]:443 307* transfer 308``` 309
