libcurl-the-guide

$Id$
                                  _   _ ____  _     
                              ___| | | |  _ \| |    
                             / __| | | | |_) | |    
                            | (__| |_| |  _ <| |___ 
                             \___|\___/|_| \_\_____|

PROGRAMMING WITH LIBCURL

About this Document

 This document will attempt to describe the general principle and some basic
 approaches to consider when programming with libcurl. The text will focus
 mainly on the C interface but might apply fairly well on other interfaces as
 well as they usually follow the C one pretty closely.

 This document will refer to 'the user' as the person writing the source code
 that uses libcurl. That would probably be you or someone in your position.
 What will be generally refered to as 'the program' will be the collected
 source code that you write that is using libcurl for transfers. The program
 is outside libcurl and libcurl is outside of the program.

 To get the more details on all options and functions described herein, please
 refer to their respective man pages.

Building

 There are many different ways to build C programs. This chapter will assume a
 unix-style build process. If you use a different build system, you can still
 read this to get general information that may apply to your environment as
 well.

  Compiling the Program

    Your compiler needs to know where the libcurl headers are
    located. Therefore you must set your compiler's include path to point to
    the directory where you installed them. The 'curl-config'[3] tool can be
    used to get this information:

        $ curl-config --cflags

  Linking the Program with libcurl

    When having compiled the program, you need to link your object files to
    create a single executable. For that to succeed, you need to link with
    libcurl and possibly also with other libraries that libcurl itself depends
    on. Like OpenSSL librararies, but even some standard OS libraries may be
    needed on the command line. To figure out which flags to use, once again
    the 'curl-config' tool comes to the rescue:

        $ curl-config --libs

  SSL or Not

    libcurl can be built and customized in many ways. One of the things that
    varies from different libraries and builds is the support for SSL-based
    transfers, like HTTPS and FTPS. If OpenSSL was detected properly at
    build-time, libcurl will be built with SSL support. To figure out if an
    installed libcurl has been built with SSL support enabled, use
    'curl-config' like this:

        $ curl-config --feature

    And if SSL is supported, the keyword 'SSL' will be written to stdout,
    possibly together with a few other features that can be on and off on
    different libcurls.


Portable Code in a Portable World

 The people behind libcurl have put a considerable effort to make libcurl work
 on a large amount of different operating systems and environments.

 You program libcurl the same way on all platforms that libcurl runs on. There
 are only very few minor considerations that differs. If you just make sure to
 write your code portable enough, you may very well create yourself a very
 portable program. libcurl shouldn't stop you from that.


Global Preparation

 The program must initialize some of the libcurl functionality globally. That
 means it should be done exactly once, no matter how many times you intend to
 use the library. Once for your program's entire life time. This is done using

    curl_global_init()

 and it takes one parameter which is a bit pattern that tells libcurl what to
 intialize. Using CURL_GLOBAL_ALL will make it initialize all known internal
 sub modules, and might be a good default option. The current two bits that
 are specified are:

  CURL_GLOBAL_WIN32 which only does anything on Windows machines. When used on
  a Windows machine, it'll make libcurl intialize the win32 socket
  stuff. Without having that initialized properly, your program cannot use
  sockets properly. You should only do this once for each application, so if
  your program already does this or of another library in use does it, you
  should not tell libcurl to do this as well.

  CURL_GLOBAL_SSL which only does anything on libcurls compiled and built
  SSL-enabled. On these systems, this will make libcurl init OpenSSL properly
  for this application. This is only needed to do once for each application so
  if your program or another library already does this, this bit should not be
  needed.

 libcurl has a default protection mechanism that detects if curl_global_init()
 hasn't been called by the time curl_easy_perform() is called and if that is
 the case, libcurl runs the function itself with a guessed bit pattern. Please
 note that depending solely on this is not considered nice nor very good.

 When the program no longer uses libcurl, it should call
 curl_global_cleanup(), which is the opposite of the init call. It will then
 do the reversed operations to cleanup the resources the curl_global_init()
 call initialized.

 Repeated calls to curl_global_init() and curl_global_cleanup() should be
 avoided. They should only be called once each.

Handle the Easy libcurl

 libcurl version 7 is oriented around the so called easy interface. All
 operations in the easy interface are prefixed with 'curl_easy'.

 Future libcurls will also offer the multi interface. More about that
 interface, what it is targeted for and how to use it is still only debated on
 the libcurl mailing list and developer web pages. Join up to discuss and
 figure out!

 To use the easy interface, you must first create yourself an easy handle. You
 need one handle for each easy session you want to perform. Basicly, you
 should use one handle for every thread you plan to use for transferring. You
 must never share the same handle in multiple threads.

 Get an easy handle with

    easyhandle = curl_easy_init();

 It returns an easy handle. Using that you proceed to the next step: setting
 up your preferred actions. A handle is just a logic entity for the upcoming
 transfer or series of transfers.

 You set properties and options for this handle using curl_easy_setopt(). They
 control how the subsequent transfer or transfers will be made. Options remain
 set in the handle until set again to something different. Alas, multiple
 requests using the same handle will use the same options.

 Many of the informationals you set in libcurl are "strings", pointers to data
 terminated with a zero byte. Keep in mind that when you set strings with
 curl_easy_setopt(), libcurl will not copy the data. It will merely point to
 the data. You MUST make sure that the data remains available for libcurl to
 use until finished or until you use the same option again to point to
 something else.

 One of the most basic properties to set in the handle is the URL. You set
 your preferred URL to transfer with CURLOPT_URL in a manner similar to:

    curl_easy_setopt(easyhandle, CURLOPT_URL, "http://curl.haxx.se/");

 Let's assume for a while that you want to receive data as the URL indentifies
 a remote resource you want to get here. Since you write a sort of application
 that needs this transfer, I assume that you would like to get the data passed
 to you directly instead of simply getting it passed to stdout. So, you write
 your own function that matches this prototype:

    size_t write_data(void *buffer, size_t size, size_t nmemb, void *userp);

 You tell libcurl to pass all data to this function by issuing a function
 similar to this:

    curl_easy_setopt(easyhandle, CURLOPT_WRITEFUNCTION, write_data);

 You can control what data your function get in the forth argument by setting
 another property:

    curl_easy_setopt(easyhandle, CURLOPT_FILE, &internal_struct);

 Using that property, you can easily pass local data between your application
 and the function that gets invoked by libcurl. libcurl itself won't touch the
 data you pass with CURLOPT_FILE.

 libcurl offers its own default internal callback that'll take care of the
 data if you don't set the callback with CURLOPT_WRITEFUNCTION. It will then
 simply output the received data to stdout. You can have the default callback
 write the data to a different file handle by passing a 'FILE *' to a file
 opened for writing with the CURLOPT_FILE option.

 Now, we need to take a step back and have a deep breath. Here's one of those
 rare platform-dependent nitpicks. Did you spot it? On some platforms[2],
 libcurl won't be able to operate on files opened by the program. Thus, if you
 use the default callback and pass in a an open file with CURLOPT_FILE, it
 will crash. You should therefore avoid this to make your program run fine
 virtually everywhere.

 There are of course many more options you can set, and we'll get back to a
 few of them later. Let's instead continue to the actual transfer:

    success = curl_easy_perform(easyhandle);

 The curl_easy_perform() will connect to the remote site, do the necessary
 commands and receive the transfer. Whenever it receives data, it calls the
 callback function we previously set. The function may get one byte at a time,
 or it may get many kilobytes at once. libcurl delivers as much as possible as
 often as possible. Your callback function should return the number of bytes
 it "took care of". If that is not the exact same amount of bytes that was
 passed to it, libcurl will abort the operation and return with an error code.

 When the transfer is complete, the function returns a return code that
 informs you if it succeeded in its mission or not. If a return code isn't
 enough for you, you can use the CURLOPT_ERRORBUFFER to point libcurl to a
 buffer of yours where it'll store a human readable error message as well.

 If you then want to transfer another file, the handle is ready to be used
 again. Mind you, it is even preferred that you re-use an existing handle if
 you intend to make another transfer. libcurl will then attempt to re-use the
 previous


When It Doesn't Work

 There will always be times when the transfer fails for some reason. You might
 have set the wrong libcurl option or misunderstood what the libcurl option
 actually does, or the remote server might return non-standard replies that
 confuse the library which then confuses your program.

 There's one golden rule when these things occur: set the CURLOPT_VERBOSE
 option to TRUE. It'll cause the library to spew out the entire protocol
 details it sends, some internal info and some received protcol data as well
 (especially when using FTP). If you're using HTTP, adding the headers in the
 received output to study is also a clever way to get a better understanding
 wht the server behaves the way it does. Include headers in the normal body
 output with CURLOPT_HEADER set TRUE.

 Of course there are bugs left. We need to get to know about them to be able
 to fix them, so we're quite dependent on your bug reports! When you do report
 suspected bugs in libcurl, please include as much details you possibly can: a
 protocol dump that CURLOPT_VERBOSE produces, library version, as much as
 possible of your code that uses libcurl, operating system name and version,
 compiler name and version etc.

 Getting some in-depth knowledge about the protocols involved is never wrong,
 and if you're trying to do funny things, you might very well understand
 libcurl and how to use it better if you study the appropriate RFC documents
 at least briefly.


Upload Data to a Remote Site

 libcurl tries to keep a protocol independent approach to most transfers, thus
 uploading to a remote FTP site is very similar to uploading data to a HTTP
 server with a PUT request.

 Of course, first you either create an easy handle or you re-use one existing
 one. Then you set the URL to operate on just like before. This is the remote
 URL, that we now will upload.

 Since we write an application, we most likely want libcurl to get the upload
 data by asking us for it. To make it do that, we set the read callback and
 the custom pointer libcurl will pass to our read callback. The read callback
 should have a prototype similar to:

    size_t function(char *bufptr, size_t size, size_t nitems, void *userp);

 Where bufptr is the pointer to a buffer we fill in with data to upload and
 size*nitems is the size of the buffer and therefore also the maximum amount
 of data we can return to libcurl in this call. The 'userp' pointer is the
 custom pointer we set to point to a struct of ours to pass private data
 between the application and the callback.

    curl_easy_setopt(easyhandle, CURLOPT_READFUNCTION, read_function);

    curl_easy_setopt(easyhandle, CURLOPT_INFILE, &filedata);

 Tell libcurl that we want to upload:

    curl_easy_setopt(easyhandle, CURLOPT_UPLOAD, TRUE);

 A few protocols won't behave properly when uploads are done without any prior
 knowledge of the expected file size. HTTP PUT is one example [1]. So, set the
 upload file size using the CURLOPT_INFILESIZE like this:

    curl_easy_setopt(easyhandle, CURLOPT_INFILESIZE, file_size);

 When you call curl_easy_perform() this time, it'll perform all the necessary
 operations and when it has invoked the upload it'll call your supplied
 callback to get the data to upload. The program should return as much data as
 possible in every invoke, as that is likely to make the upload perform as
 fast as possible. The callback should return the number of bytes it wrote in
 the buffer. Returning 0 will signal the end of the upload.


Passwords

 Many protocols use or even require that user name and password are provided
 to be able to download or upload the data of your choice. libcurl offers
 several ways to specify them.

 Most protocols support that you specify the name and password in the URL
 itself. libcurl will detect this and use them accordingly. This is written
 like this:

        protocol://user:password@example.com/path/

 If you need any odd letters in your user name or password, you should enter
 them URL encoded, as %XX where XX is a two-digit hexadecimal number.

 libcurl also provides options to set various passwords. The user name and
 password as shown embedded in the URL can instead get set with the
 CURLOPT_USERPWD option. The argument passed to libcurl should be a char * to
 a string in the format "user:password:". In a manner like this:

        curl_easy_setopt(easyhandle, CURLOPT_USERPWD, "myname:thesecret");

 Another case where name and password might be needed at times, is for those
 users who need to athenticate themselves to a proxy they use. libcurl offers
 another option for this, the CURLOPT_PROXYUSERPWD. It is used quite similar
 to the CURLOPT_USERPWD option like this:

        curl_easy_setopt(easyhandle, CURLOPT_PROXYUSERPWD, "myname:thesecret");
 
 There's a long time unix "standard" way of storing ftp user names and
 passwords, namely in the $HOME/.netrc file. The file should be made private
 so that only the user may read it (see also the "Security Considerations"
 chapter), as it might contain the password in plain text. libcurl has the
 ability to use this file to figure out what set of user name and password to
 use for a particular host. As an extension to the normal functionality,
 libcurl also supports this file for non-FTP protocols such as HTTP. To make
 curl use this file, use the CURLOPT_NETRC option:

    curl_easy_setopt(easyhandle, CURLOPT_NETRC, TRUE);

 And a very basic example of how such a .netrc file may look like:

    machine myhost.mydomain.com
    login userlogin
    password secretword

 All these examples have been cases where the password has been optional, or
 at least you could leave it out and have libcurl attempt to do its job
 without it. There are times when the password isn't optional, like when
 you're using an SSL private key for secure transfers.

 You can in this situation either pass a password to libcurl to use to unlock
 the private key, or you can let libcurl prompt the user for it. If you prefer
 to ask the user, then you can provide your own callback function that will be
 called when libcurl wants the password. That way, you can control how the
 question will appear to the user.

 To pass the known private key password to libcurl:

    curl_easy_setopt(easyhandle, CURLOPT_SSLKEYPASSWD, "keypassword");

 To make a password callback:

    int enter_passwd(void *ourp, const char *prompt, char *buffer, int len);
    curl_easy_setopt(easyhandle, CURLOPT_PASSWDFUNCTION, enter_passwd);


HTTP POSTing

 We get many questions regarding how to issue HTTP POSTs with libcurl the
 proper way. This chapter will thus include examples using both different
 versions of HTTP POST that libcurl supports.

 The first version is the simple POST, the most common version, that most HTML
 pages using the <form> tag uses. We provide a pointer to the data and tell
 libcurl to post it all to the remote site:

    char *data="name=daniel&project=curl";
    curl_easy_setopt(easyhandle, CURLOPT_POSTFIELDS, data);
    curl_easy_setopt(easyhandle, CURLOPT_URL, "http://posthere.com/");

    curl_easy_perform(easyhandle); /* post away! */

 Simple enough, huh? Since you set the POST options with the
 CURLOPT_POSTFIELDS, this automaticly switches the handle to use POST in the
 upcoming request.

 Ok, so what if you want to post binary data that also requires you to set the
 Content-Type: header of the post? Well, binary posts prevents libcurl from
 being able to do strlen() on the data to figure out the size, so therefore we
 must tell libcurl the size of the post data. Setting headers in libcurl
 requests are done in a generic way, by building a list of our own headers and
 then passing that list to libcurl.

    struct curl_slist *headers=NULL;
    headers = curl_slist_append(headers, "Content-Type: text/xml");

    /* post binary data */
    curl_easy_setopt(easyhandle, CURLOPT_POSTFIELD, binaryptr);

    /* set the size of the postfields data */
    curl_easy_setopt(easyhandle, CURLOPT_POSTFIELDSIZE, 23);

    /* pass our list of custom made headers */
    curl_easy_setopt(easyhandle, CURLOPT_HTTPHEADER, headers);

    curl_easy_perform(easyhandle); /* post away! */

    curl_slist_free_all(headers); /* free the header list */

 While the simple examples above cover the majority of all cases where HTTP
 POST operations are required, they don't do multipart formposts. Multipart
 formposts were introduced as a better way to post (possibly large) binary
 data and was first documented in the RFC1867. They're called multipart
 because they're built by a chain of parts, each being a single unit. Each
 part has its own name and contents. You can in fact create and post a
 multipart formpost with the regular libcurl POST support described above, but
 that would require that you build a formpost yourself and provide to
 libcurl. To make that easier, libcurl provides curl_formadd(). Using this
 function, you add parts to the form. When you're done adding parts, you post
 the whole form.

 The following example sets two simple text parts with plain textual contents,
 and then a file with binary contents and upload the whole thing.

    struct HttpPost *post=NULL;
    struct HttpPost *last=NULL;
    curl_formadd(&post, &last,
                 CURLFORM_COPYNAME, "name",
                 CURLFORM_COPYCONTENTS, "daniel", CURLFORM_END);
    curl_formadd(&post, &last,
                 CURLFORM_COPYNAME, "project",
                 CURLFORM_COPYCONTENTS, "curl", CURLFORM_END);
    curl_formadd(&post, &last,
                 CURLFORM_COPYNAME, "logotype-image",
                 CURLFORM_FILECONTENT, "curl.png", CURLFORM_END);

    /* Set the form info */
    curl_easy_setopt(easyhandle, CURLOPT_HTTPPOST, post);

    curl_easy_perform(easyhandle); /* post away! */

    /* free the post data again */
    curl_formfree(post);

 Multipart formposts are chains of parts using MIME-style separators and
 headers. It means that each one of these separate parts get a few headers set
 that describe the individual content-type, size etc. To enable your
 application to handicraft this formpost even more, libcurl allows you to
 supply your own set of custom headers to such an individual form part. You
 can of course supply headers to as many parts you like, but this little
 example will show how you set headers to one specific part when you add that
 to the post handle:

    struct curl_slist *headers=NULL;
    headers = curl_slist_append(headers, "Content-Type: text/xml");

    curl_formadd(&post, &last,
                 CURLFORM_COPYNAME, "logotype-image",
                 CURLFORM_FILECONTENT, "curl.xml",
                 CURLFORM_CONTENTHEADER, headers,
                 CURLFORM_END);

    curl_easy_perform(easyhandle); /* post away! */

    curl_formfree(post); /* free post */
    curl_slist_free_all(post); /* free custom header list */

 Since all options on an easyhandle are "sticky", they remain the same until
 changed even if you do call curl_easy_perform(), you may need to tell curl to
 go back to a plain GET request if you intend to do such a one as your next
 request. You force an easyhandle to back to GET by using the CURLOPT_HTTPGET
 option:

    curl_easy_setopt(easyhandle, CURLOPT_HTTPGET, TRUE);

 Just setting CURLOPT_POSTFIELDS to "" or NULL will *not* stop libcurl from
 doing a POST. It will just make it POST without any data to send!


Showing Progress

 For historical and traditional reasons, libcurl has a built-in progress meter
 that can be switched on and then makes it presents a progress meter in your
 terminal.

 Switch on the progress meter by, oddly enough, set CURLOPT_NOPROGRESS to
 FALSE. This option is set to TRUE by default.

 For most applications however, the built-in progress meter is useless and
 what instead is interesting is the ability to specify a progress
 callback. The function pointer you pass to libcurl will then be called on
 irregular intervals with information about the current transfer.

 Set the progress callback by using CURLOPT_PROGRESSFUNCTION. And pass a
 pointer to a function that matches this prototype:

        int progress_callback(void *clientp,
                              double dltotal,
                              double dlnow,
                              double ultotal,
                              double ulnow);

 If any of the input arguments is unknown, a 0 will be passed. The first
 argument, the 'clientp' is the pointer you pass to libcurl with
 CURLOPT_PROGRESSDATA. libcurl won't touch it.


libcurl with C++

 There's basicly only one thing to keep in mind when using C++ instead of C
 when interfacing libcurl:

    "The Callbacks Must Be Plain C"

 So if you want a write callback set in libcurl, you should put it within
 'extern'. Similar to this:

     extern "C" {
       size_t write_data(void *ptr, size_t size, size_t nmemb,
                         void *ourpointer)
       {
         /* do what you want with the data */
       }
    }

 This will of course effectively turn the callback code into C. There won't be
 any "this" pointer available etc.


Proxies

 What "proxy" means according to Merriam-Webster: "a person authorized to act
 for another" but also "the agency, function, or office of a deputy who acts
 as a substitute for another".

 Proxies are exceedingly common these days. Companies often only offer
 internet access to employees through their HTTP proxies. Network clients or
 user-agents ask the proxy for docuements, the proxy does the actual request
 and then it returns them.

 libcurl has full support for HTTP proxies, so when a given URL is wanted,
 libcurl will ask the proxy for it instead of trying to connect to the actual
 host identified in the URL.

 The fact that the proxy is a HTTP proxy puts certain restrictions on what can
 actually happen. A requested URL that might not be a HTTP URL will be still
 be passed to the HTTP proxy to deliver back to libcurl. This happens
 transparantly, and an application may not need to know. I say "may", because
 at times it is very important to understand that all operations over a HTTP
 proxy is using the HTTP protocol. For example, you can't invoke your own
 custom FTP commands or even proper FTP directory listings.

  Proxy Options

    To tell libcurl to use a proxy at a given port number:

       curl_easy_setopt(easyhandle, CURLOPT_PROXY, "proxy-host.com:8080");

    Some proxies require user authentication before allowing a request, and
    you pass that information similar to this:

       curl_easy_setopt(easyhandle, CURLOPT_PROXYUSERPWD, "user:password");

    If you want to, you can specify the host name only in the CURLOPT_PROXY
    option, and set the port number separately with CURLOPT_PROXYPORT.

  Environment Variables

    libcurl automaticly checks and uses a set of environment variables to know
    what proxies to use for certain protocols. The names of the variables are
    following an ancient de facto standard and are built up as
    "[protocol]_proxy" (note the lower casing). Which makes the variable
    'http_proxy' checked for a name of a proxy to use when the input URL is
    HTTP. Following the same rule, the variable named 'ftp_proxy' is checked
    for FTP URLs. Again, the proxies are always HTTP proxies, the different
    names of the variables simply allows different HTTP proxies to be used.

    The proxy environment variable contents should be in the format
    "[protocol://]machine[:port]". Where the protocol:// part is simply
    ignored if present (so http://proxy and bluerk://proxy will do the same)
    and the optional port number specifies on which port the proxy operates on
    the host. If not specified, the internal default port number will be used
    and that is most likely *not* the one you would like it to be.

    There are two special environment variables. 'all_proxy' is what sets
    proxy for any URL in case the protocol specific variable wasn't set, and
    'no_proxy' defines a list of hosts that should not use a proxy even though
    a variable may say so. If 'no_proxy' is a plain asterisk ("*") it matches
    all hosts.

  SSL and Proxies

    SSL is for secure point-to-point connections. This involves strong
    encryption and similar things, which effectivly makes it impossible for a
    proxy to operate as a "man in between" which the proxy's task is, as
    previously discussed. Instead, the only way to have SSL work over a HTTP
    proxy is to ask the proxy to tunnel trough everything without being able
    to check or fiddle with the traffic.

    Opening an SSL connection over a HTTP proxy is therefor a matter of asking
    the proxy for a straight connection to the target host on a specified
    port. This is made with the HTTP request CONNECT. ("please mr proxy,
    connect me to that remote host").

    Because of the nature of this operation, where the proxy has no idea what
    kind of data that is passed in and out through this tunnel, this breaks
    some of the very few advantages that come from using a proxy, such as
    caching.  Many organizations prevent this kind of tunneling to other
    destination port numbers than 443 (which is the default HTTPS port
    number).

  Tunneling Through Proxy

    As explained above, tunneling is required for SSL to work and often even
    restricted to the operation intended for SSL; HTTPS.

    This is however not the only time proxy-tunneling might offer benefits to
    you or your application.

    As tunneling opens a direct connection from your application to the remote
    machine, it suddenly also re-introduces the ability to do non-HTTP
    operations over a HTTP proxy. You can in fact use things such as FTP
    upload or FTP custom commands this way.

    Again, this is often prevented by the adminstrators of proxies and is
    rarely allowed.

    Tell libcurl to use proxy tunneling like this:

       curl_easy_setopt(easyhandle, CURLOPT_HTTPPROXYTUNNEL, TRUE);

    In fact, there might even be times when you want to do plain HTTP
    operations using a tunnel like this, as it then enables you to operate on
    the remote server instead of asking the proxy to do so. libcurl will not
    stand in the way for such innovative actions either!

  Proxy Auto-Config

    Netscape first came up with this. It is basicly a web page (usually using
    a .pac extension) with a javascript that when executed by the browser with
    the requested URL as input, returns information to the browser on how to
    connect to the URL. The returned information might be "DIRECT" (which
    means no proxy should be used), "PROXY host:port" (to tell the browser
    where the proxy for this particular URL is) or "SOCKS host:port" (to
    direct the brower to a SOCKS proxy).

    libcurl has no means to interpret or evaluate javascript and thus it
    doesn't support this. If you get yourself in a position where you face
    this nasty invention, the following advice have been mentioned and used in
    the past:

    - Depending on the javascript complexity, write up a script that
      translates it to another language and execute that.

    - Read the javascript code and rewrite the same logic in another language.

    - Implement a javascript interpreted, people have successfully used the
      Mozilla javascript engine in the past.

    - Ask your admins to stop this, for a static proxy setup or similar.


Persistancy Is The Way to Happiness

 Re-cycling the same easy handle several times when doing multiple requests is
 the way to go.

 After each single curl_easy_perform() operation, libcurl will keep the
 connection alive and open. A subsequent request using the same easy handle to
 the same host might just be able to use the already open connection! This
 reduces network impact a lot.

 Even if the connection is dropped, all connections involving SSL to the same
 host again, will benefit from libcurl's session ID cache that drasticly
 reduces re-connection time.

 FTP connections that are kept alive saves a lot of time, as the command-
 response roundtrips are skipped, and also you don't risk getting blocked
 without permission to login again like on many FTP servers only allowing N
 persons to be logged in at the same time.

 libcurl caches DNS name resolving results, to make lookups of a previously
 looked up name a lot faster.

 Other interesting details that improve performance for subsequent requests
 may also be added in the future.

 Each easy handle will attempt to keep the last few connections alive for a
 while in case they are to be used again. You can set the size of this "cache"
 with the CURLOPT_MAXCONNECTS option. Default is 5. It is very seldom any
 point in changing this value, and if you think of changing this it is often
 just a matter of thinking again.

 When the connection cache gets filled, libcurl must close an existing
 connection in order to get room for the new one. To know which connection to
 close, libcurl uses a "close policy" that you can affect with the
 CURLOPT_CLOSEPOLICY option. There's only two polices implemented as of this
 writing (libcurl 7.9.4) and they are:

  CURLCLOSEPOLICY_LEAST_RECENTLY_USED simply close the one that hasn't been
  used for the longest time. This is the default behavior.

  CURLCLOSEPOLICY_OLDEST closes the oldest connection, the one that was
  createst the longest time ago.

 There are, or at least were, plans to support a close policy that would call
 a user-specified callback to let the user be able to decide which connection
 to dump when this is necessary and therefor is the CURLOPT_CLOSEFUNCTION an
 existing option still today. Nothing ever uses this though and this will not
 be used within the forseeable future either.

 To force your upcoming request to not use an already existing connection (it
 will even close one first if there happens to be one alive to the same host
 you're about to operate on), you can do that by setting CURLOPT_FRESH_CONNECT
 to TRUE. In a similar spirit, you can also forbid the upcoming request to be
 "lying" around and possibly get re-used after the request by setting
 CURLOPT_FORBID_REUSE to TRUE.


Customizing Operations

 There is an ongoing development today where more and more protocols are built
 upon HTTP for transport. This has obvious benefits as HTTP is a tested and
 reliable protocol that is widely deployed and have excellent proxy-support.

 When you use one of these protocols, and even when doing other kinds of
 programming you may need to change the traditional HTTP (or FTP or...)
 manners. You may need to change words, headers or various data.

 libcurl is your friend here too.

 If just changing the actual HTTP request keyword is what you want, like when
 GET, HEAD or POST is not good enough for you, CURLOPT_CUSTOMREQUEST is there
 for you. It is very simple to use:

    curl_easy_setopt(easyhandle, CURLOPT_CUSTOMREQUEST, "MYOWNRUQUEST");

 When using the custom request, you change the request keyword of the actual
 request you are performing. Thus, by default you make GET request but you can
 also make a POST operation (as described before) and then replace the POST
 keyword if you want to. You're the boss.

 HTTP-like protocols pass a series of headers to the server when doing the
 request, and you're free to pass any amount of extra headers that you think
 fit. Adding headers are this easy:

    struct curl_slist *headers;

    headers = curl_slist_append(headers, "Hey-server-hey: how are you?");
    headers = curl_slist_append(headers, "X-silly-content: yes");

    /* pass our list of custom made headers */
    curl_easy_setopt(easyhandle, CURLOPT_HTTPHEADER, headers);

    curl_easy_perform(easyhandle); /* transfer http */

    curl_slist_free_all(headers); /* free the header list */

 ... and if you think some of the internally generated headers, such as
 User-Agent:, Accept: or Host: don't contain the data you want them to
 contain, you can replace them by simply setting them too:

    headers = curl_slist_append(headers, "User-Agent: 007");
    headers = curl_slist_append(headers, "Host: munged.host.line");

 If you replace an existing header with one with no contents, you will prevent
 the header from being sent. Like if you want to completely prevent the
 "Accept:" header to be sent, you can disable it with code similar to this:

    headers = curl_slist_append(headers, "Accept:");

 Both replacing and cancelling internal headers should be done with careful
 consideration and you should be aware that you may violate the HTTP protocol
 when doing so.

 There's only one aspect left in the HTTP requests that we haven't yet
 mentioned how to modify: the version field. All HTTP requests includes the
 version number to tell the server which version we support. libcurl speak
 HTTP 1.1 by default. Some very old servers don't like getting 1.1-requests
 and when dealing with stubborn old things like that, you can tell libcurl to
 use 1.0 instead by doing something like this:

    curl_easy_setopt(easyhandle, CURLOPT_HTTP_VERSION, CURLHTTP_VERSION_1_0);

 Not all protocols are HTTP-like, and thus the above may not help you when you
 want to make for example your FTP transfers to behave differently.

 Sending custom commands to a FTP server means that you need to send the
 comands exactly as the FTP server expects them (RFC959 is a good guide here),
 and you can only use commands that work on the control-connection alone. All
 kinds of commands that requires data interchange and thus needs a
 data-connection must be left to libcurl's own judgement. Also be aware that
 libcurl will do its very best to change directory to the target directory
 before doing any transfer, so if you change directory (with CWD or similar)
 you might confuse libcurl and then it might not attempt to transfer the file
 in the correct remote directory.

 A little example that deletes a given file before an operation:

    headers = curl_slist_append(headers, "DELE file-to-remove");

    /* pass the list of custom commands to the handle */
    curl_easy_setopt(easyhandle, CURLOPT_QUOTE, headers);

    curl_easy_perform(easyhandle); /* transfer ftp data! */

    curl_slist_free_all(headers); /* free the header list */

 If you would instead want this operation (or chain of operations) to happen
 _after_ the data transfer took place the option to curl_easy_setopt() would
 instead be called CURLOPT_POSTQUOTE and used the exact same way.

 The custom FTP command will be issued to the server in the same order they
 are added to the list, and if a command gets an error code returned back from
 the server, no more commands will be issued and libcurl will bail out with an
 error code (CURLE_FTP_QUOTE_ERROR). Note that if you use CURLOPT_QUOTE to
 send commands before a transfer, no transfer will actually take place when a
 quote command has failed.

 If you set the CURLOPT_HEADER to true, you will tell libcurl to get
 information about the target file and output "headers" about it. The headers
 will be in "HTTP-style", looking like they do in HTTP.

 The option to enable headers or to run custom FTP commands may be useful to
 combine with CURLOPT_NOBODY. If this option is set, no actual file content
 transfer will be performed.


Cookies Without Chocolate Chips

 In the HTTP sense, a cookie is a name with an associated value. A server
 sends the name and value to the client, and expects it to get sent back on
 every subsequent request to the server that matches the particular conditions
 set. The conditions include that the domain name and path match and that the
 cookie hasn't become too old.

 In real-world cases, servers send new cookies to replace existing one to
 update them. Server use cookies to "track" users and to keep "sessions".

 Cookies are sent from server to clients with the header Set-Cookie: and
 they're sent from clients to servers with the Cookie: header.

 To just send whatever cookie you want to a server, you can use CURLOPT_COOKIE
 to set a cookie string like this:

    curl_easy_setopt(easyhandle, CURLOPT_COOKIE, "name1=var1; name2=var2;");

 In many cases, that is not enough. You might want to dynamicly save whatever
 cookies the remote server passes to you, and make sure those cookies are then
 use accordingly on later requests.

 One way to do this, is to save all headers you receive in a plain file and
 when you make a request, you tell libcurl to read the previous headers to
 figure out which cookies to use. Set header file to read cookies from with
 CURLOPT_COOKIEFILE.

 The CURLOPT_COOKIEFILE option also automaticly enables the cookie parser in
 libcurl. Until the cookie parser is enabled, libcurl will not parse or
 understand incoming cookies and they will just be ignored. However, when the
 parser is enabled the cookies will be understood and the cookies will be kept
 in memory and used properly in subsequent requests when the same handle is
 used. Many times this is enough, and you may not have to save the cookies to
 disk at all. Note that the file you specify to CURLOPT_COOKIEFILE doesn't
 have to exist to enable the parser, so a common way to just enable the parser
 and not read able might be to use a file name you know doesn't exist.

 If you rather use existing cookies that you've previously received with your
 Netscape or Mozilla browsers, you can make libcurl use that cookie file as
 input. The CURLOPT_COOKIEFILE is used for that too, as libcurl will
 automaticly find out what kind of file it is and act accordingly.

 The perhaps most advanced cookie operation libcurl offers, is saving the
 entire internal cookie state back into a Netscape/Mozilla formatted cookie
 file. We call that the cookie-jar. When you set a file name with
 CURLOPT_COOKIEJAR, that file name will be created and all received cookies
 will be stored in it when curl_easy_cleanup() is called. This enabled cookies
 to get passed on properly between multiple handles without any information
 getting lost.


Headers Equal Fun

 [ use the header callback for HTTP, FTP etc ]


Post Transfer Information

 [ curl_easy_getinfo ]


Security Considerations

 [ ps output, netrc plain text, plain text protocols / base64 ]


SSL, Certificates and Other Tricks

 [ seeding, passwords, keys, certificates, ENGINE, ca certs ]


Future

 [ multi interface, sharing between handles, mutexes, pipelining ]


-----
Footnotes:

[1] = HTTP PUT without knowing the size prior to transfer is indeed possible,
      but libcurl does not support the chunked transfers on uploading that is
      necessary for this feature to work. We'd gratefully appreciate patches
      that bring this functionality...

[2] = This happens on Windows machines when libcurl is built and used as a
      DLL. However, you can still do this on Windows if you link with a static
      library.

[3] = The curl-config tool is generated at build-time (on unix-like systems)
      and should be installed with the 'make install' or similar instruction
      that installs the library, header files, man pages etc.