curl/docs/CLIENT-READERS.md

curl client readers

Client readers is a design in the internals of libcurl, not visible in its public API. They were started in curl v8.7.0. This document describes the concepts, its high level implementation and the motivations.

Naming

libcurl operates between clients and servers. A client is the application using libcurl, like the command line tool curl itself. Data to be uploaded to a server is read from the client and sent to the server, the servers response is received by libcurl and then written to the client.

With this naming established, client readers are concerned with providing data from the application to the server. Applications register callbacks via CURLOPT_READFUNCTION, data via CURLOPT_POSTFIELDS and other options to be used by libcurl when the request is send.

Invoking

The transfer loop that sends and receives, is using Curl_client_read() to get more data to send for a transfer. If no specific reader has been installed yet, the default one that uses CURLOPT_READFUNCTION is added. The prototype is

CURLcode Curl_client_read(struct Curl_easy *data, char *buf, size_t blen,
                          size_t *nread, bool *eos);

The arguments are the transfer to read for, a buffer to hold the read data, its length, the actual number of bytes placed into the buffer and the eos (end of stream) flag indicating that no more data is available. The eos flag may be set for a read amount, if that amount was the last. That way curl can avoid to read an additional time.

The implementation of Curl_client_read() uses a chain of client reader instances to get the data. This is similar to the design of client writers. The chain of readers allows processing of the data to send.

The definition of a reader is:

struct Curl_crtype {
  const char *name;        /* writer name. */
  CURLcode (*do_init)(struct Curl_easy *data, struct Curl_creader *writer);
  CURLcode (*do_read)(struct Curl_easy *data, struct Curl_creader *reader,
                      char *buf, size_t blen, size_t *nread, bool *eos);
  void (*do_close)(struct Curl_easy *data, struct Curl_creader *reader);
  bool (*needs_rewind)(struct Curl_easy *data, struct Curl_creader *reader);
  curl_off_t (*total_length)(struct Curl_easy *data,
                             struct Curl_creader *reader);
  CURLcode (*resume_from)(struct Curl_easy *data,
                          struct Curl_creader *reader, curl_off_t offset);
  CURLcode (*rewind)(struct Curl_easy *data, struct Curl_creader *reader);
};

struct Curl_creader {
  const struct Curl_crtype *crt;  /* type implementation */
  struct Curl_creader *next;  /* Downstream reader. */
  Curl_creader_phase phase; /* phase at which it operates */
};

Curl_creader is a reader instance with a next pointer to form the chain. It as a type crt which provides the implementation. The main callback is do_read() which provides the data to the caller. The others are for setup and tear down. needs_rewind() is explained further below.

Phases and Ordering

Since client readers may transform the data being read through the chain, the order in which they are called is relevant for the outcome. When a reader is created, it gets the phase property in which it operates. Reader phases are defined like:

typedef enum {
  CURL_CR_NET,  /* data send to the network (connection filters) */
  CURL_CR_TRANSFER_ENCODE, /* add transfer-encodings */
  CURL_CR_PROTOCOL, /* before transfer, but after content decoding */
  CURL_CR_CONTENT_ENCODE, /* add content-encodings */
  CURL_CR_CLIENT  /* data read from client */
} Curl_creader_phase;

If a reader for phase PROTOCOL is added to the chain, it is always added after any NET or TRANSFER_ENCODE readers and before and CONTENT_ENCODE and CLIENT readers. If there is already a reader for the same phase, the new reader is added before the existing one(s).

Example: chunked reader

In http_chunks.c a client reader for chunked uploads is implemented. This one operates at phase CURL_CR_TRANSFER_ENCODE. Any data coming from the reader "below" has the HTTP/1.1 chunk handling applied and returned to the caller.

When this reader sees an eos from below, it generates the terminal chunk, adding trailers if provided by the application. When that last chunk is fully returned, it also sets eos to the caller.

Example: lineconv reader

In sendf.c a client reader that does line-end conversions is implemented. It operates at CURL_CR_CONTENT_ENCODE and converts any "\n" to "\r\n". This is used for FTP ASCII uploads or when the general crlf options has been set.

Example: null reader

Implemented in sendf.c for phase CURL_CR_CLIENT, this reader has the simple job of providing transfer bytes of length 0 to the caller, immediately indicating an eos. This reader is installed by HTTP for all GET/HEAD requests and when authentication is being negotiated.

Example: buf reader

Implemented in sendf.c for phase CURL_CR_CLIENT, this reader get a buffer pointer and a length and provides exactly these bytes. This one is used in HTTP for sending postfields provided by the application.

Request retries

Sometimes it is necessary to send a request with client data again. Transfer handling can inquire via Curl_client_read_needs_rewind() if a rewind (e.g. a reset of the client data) is necessary. This asks all installed readers if they need it and give FALSE of none does.

Upload Size

Many protocols need to know the amount of bytes delivered by the client readers in advance. They may invoke Curl_creader_total_length(data) to retrieve that. However, not all reader chains know the exact value beforehand. In that case, the call returns -1 for "unknown".

Even if the length of the "raw" data is known, the length that is send may not. Example: with option --crlf the uploaded content undergoes line-end conversion. The line converting reader does not know in advance how many newlines it may encounter. Therefore it must return -1 for any positive raw content length.

In HTTP, once the correct client readers are installed, the protocol asks the readers for the total length. If that is known, it can set Content-Length: accordingly. If not, it may choose to add an HTTP "chunked" reader.

In addition, there is Curl_creader_client_length(data) which gives the total length as reported by the reader in phase CURL_CR_CLIENT without asking other readers that may transform the raw data. This is useful in estimating the size of an upload. The HTTP protocol uses this to determine if Expect: 100-continue shall be done.

Resuming

Uploads can start at a specific offset, if so requested. The "resume from" that offset. This applies to the reader in phase CURL_CR_CLIENT that delivers the "raw" content. Resumption can fail if the installed reader does not support it or if the offset is too large.

The total length reported by the reader changes when resuming. Example: resuming an upload of 100 bytes by 25 reports a total length of 75 afterwards.

If resume_from() is invoked twice, it is additive. There is currently no way to undo a resume.

Rewinding

When a request is retried, installed client readers are discarded and replaced by new ones. This works only if the new readers upload the same data. For many readers, this is not an issue. The "null" reader always does the same. Also the buf reader, initialized with the same buffer, does this.

Readers operating on callbacks to the application need to "rewind" the underlying content. For example, when reading from a FILE*, the reader needs to fseek() to the beginning. The following methods are used:

1. Curl_creader_needs_rewind(data): tells if a rewind is necessary, given the current state of the reader chain. If nothing really has been read so far, this returns FALSE.
2. Curl_creader_will_rewind(data): tells if the reader chain rewinds at the start of the next request.
3. Curl_creader_set_rewind(data, TRUE): marks the reader chain for rewinding at the start of the next request.
4. Curl_client_start(data): tells the readers that a new request starts and they need to rewind if requested.

Summary and Outlook

By adding the client reader interface, any protocol can control how/if it wants the curl transfer to send bytes for a request. The transfer loop becomes then blissfully ignorant of the specifics.

The protocols on the other hand no longer have to care to package data most efficiently. At any time, should more data be needed, it can be read from the client. This is used when sending HTTP requests headers to add as much request body data to the initial sending as there is room for.

Future enhancements based on the client readers:

• expect-100 handling: place that into a HTTP specific reader at CURL_CR_PROTOCOL and eliminate the checks in the generic transfer parts.
• eos forwarding: transfer should forward an eos flag to the connection filters. Filters like HTTP/2 and HTTP/3 can make use of that, terminating streams early. This would also eliminate length checks in stream handling.