Plasma GitLab Archive
Projects Blog Knowledge


Netclient tutorial

Here we give some recipes how to submit HTTP and FTP requests to servers. There is a quite simple interface representing the hierarchy of remote files, which is recommended for all occasional uses. If performance is important, however, the protocol-specific interfaces will give you more options.

Accessing HTTP and FTP as filesystem

Let's start with an example: We want to get the public files

For HTTP we need a filesystem object accessing the server:

let fs1 = Http_fs.http_fs ""

The same for FTP looks like:

let fs2 = Ftp_fs.ftp_fs ""

The objects fs1 and fs2 provide now a number of methods for accessing files. These do not only cover downloads, but also listing directories, writing files, renaming files, and a number of further operations. The commonly available methods are those of Netfs.stream_fs. The incarnations of this interface for concrete protocols usually define more methods. It is guaranteed that you can coerce the types to Netfs.stream_fs, though:

let my_filesystems =
  [ (fs1 :> Netfs.stream_fs); (fs2 :> Netfs.stream_fs) ]

The method we use here is read:

method read : read_flags -> string -> Netchannels.in_obj_channel

All access methods take a list of flags as first argument. For example, a possible flag here is `Binary switching to binary mode for the protcols where it makes a difference (like FTP).

The second argument is the file path, using slashes as separators, and always starting with a slash. The path is appended to the base URL given when creating the fs1 and fs2 objects. Note that the path must not contain any URL-specific encodings like "%xx".

We get a an Netchannels.in_obj_channel back we can read the data from:

let c1 = fs1 # read [] "/bar"
let s1 = Netchannels.string_of_in_obj_channel c1
let () = c1 # close_in()

let c2 = fs2 # read [`Binary] "/baz"
let s2 = Netchannels.string_of_in_obj_channel c2
let () = c2 # close_in()

It depends very much on the implementation what actually happens:

  • HTTP has built-in resilience against errors. Downloads are tried again if errors occur.
  • HTTP follows redirects automatically, perhaps pointing to a different server.
  • FTP does not retry after errors.
  • The possible error codes can be different. For example, FTP does not distinguish between "access denied" and "file not found".
You should keep in mind that the protocol differences do not go away, just because we are mapping the protocols to a common interface here.

By default, the downloaded data are cached in a temporary file. Some implementations support the streaming mode to avoid that (like HTTP), and you are directly connected with the reading socket when reading from the returned in_obj_channel. Pass `Streaming as flag to read to enable this. In streaming mode, however, neither retries nor redirects are possible.

Other access methods

An overview:

  • write works very much like read, only that you get a Netchannels.out_obj_channel back. The network write operation normally starts first when this channel is closed, and the so-far cached data are uploaded to the server. For HTTP there is also a streaming mode. The write operation takes also flags that look like normal open flags, i.e. whether you want to create a file, truncate a file, or ensure the unique creation. Not all protocols support every combination, though. For HTTP a write is translated to sending a PUT method to the server.
  • readdir reads the names of a directory. For FTP this is clearly an NLST command. For HTTP the implementation just extracts the names from the hyperlinks contained in the page - this works well for applying readdir to automatically generated file indexes.
  • remove translates to the DELETE method for HTTP. This method is defined in the HTTP standard, but usually not available on servers, though.
  • size gets the size of a file. This may work for HTTP or may not - depending on whether the server knows the size (which is often not the case for dynamically generated content). For FTP there is the SIZE command. However, this is a later addition to the protocol, and may not be available on ancient servers.
  • test and test_list allow it to test properties of files (existence, type, non-empty, accessibility). This is only partially implemented for HTTP and FTP.
The following operations are only applicable to FTP:

  • rename
  • mkdir
  • rmdir
There are also operations that do not make sense here:

  • symlink
  • readlink
  • copy

Other filesystem implementations

There is a full implementation of Netfs.stream_fs for accessing local files: Netfs.local_fs. There are more definitions inside and outside Ocamlnet, see Other impementations of stream_fs for a list. It also mentions a WebDAV implementation extending the HTTP definition explained here, and which covers a larger set of access operations.

Configuring HTTP and FTP file accesses

When creating the access object, one can set a callback that allows almost arbitrary configurations:

let fs1 = 
     ~config_pipeline:(fun p -> ...)

let fs2 = 
     ~config_client:(fun c -> ...)

Here, p and c are the underlying protocol implementations.

Authentication for HTTP

Do it like this in the config_pipeline callback:

  let user = "user" in
  let password = "secret" in
  let realm = "the realm string" in
  let domain = [ " " ] in
  let keys = new Http_client.key_ring() in
  keys # add_key (Http_client.key ~user ~password ~realm ~domain);
  let ah = new Http_client.unified_auth_handler keys in
  p # add_auth_handler ah

This works for both "basic" and "digest" authentication.

Authentication for FTP

This is not done in the config_client callback, but directly when creating the filesystem object. The user string is always taken from the URL (as normally the accessed file space depends on the user). Passwords and account names (if needed) are supplied by callbacks:

let fs2 = 
      ~get_password:(fun () -> "secret")
      ~get_account:(fun () -> "account")

Configure a web proxy for HTTP

Do it like this in the config_pipeline callback:

  p # set_proxy "" 8080;
  p # set_proxy_auth "user" "secret";
  p # avoid_proxy_for [ ""; "localhost" ]

Or you can just import this data from the environment variables "http_proxy" and "no_proxy":

  p # set_proxy_from_environment()

Configure a web proxy for FTP

Web proxies often also support FTP URLs, but only for a limited set of operations (often only read works).

Note that you have to use Http_fs to use this feature, not Ftp_fs:

let fs1 = 
     ~config_pipeline:(fun p ->
        p # set_proxy "" 8080;
        p # set_proxy_auth "user" "secret";

In this configuration, the web proxy is contacted via HTTP, and the proxy talks FTP with the content server.

If you do not configure the proxy, any accesses will fail (no transport error).

Configure a SOCKS proxy for HTTP

This is an alternative to a web proxy. Do it like this in the config_pipeline callback:

  p # set_socks5_proxy "" 1080

Configure a SOCKS proxy for FTP

Do it like this in the config_client callback:

  c # set_socks5_proxy "" 1080

The current implementation is limited to file transfers in passive mode, though. This is nowadays not a problem anymore, because almost all FTP servers support it.

Configure HTTPS

Support for TLS (SSL) is not available by default. Ocamlnet must be compiled with support for TLS, and a certain configuration must be applied to the HTTP pipeline.

See the Https_client module for a recipe (this module is part of equeue-ssl).


The Netglob module can be used to interpret wildcards in filenames. An example:

let files =
    ~fsys:(Netglob.of_stream_fs (fs2 :> Netfs.stream_fs))
    (`String "/dir/*.gif")

This would return paths to all gif files in /dir on the FTP server fs2.

Caveat: Globbing works only well if the server provides the operations for recognizing directories. Most FTP servers don't - only the recently (1) added MLST command allows it to safely recognize directories.

(1) recently = many years ago, but existing FTP deployments seem only to be very slowly upgraded.

Test whether an FTP server supports MLST: There must be a line for MLST in the output for the FEAT command, like in

$ ftp localhost
Connected to localhost.
220---------- Welcome to Pure-FTPd [privsep] [TLS] ----------
ftp> quote feat
211-Extensions supported:
 MLST type*;size*;sizd*;modify*;UNIX.mode*;UNIX.uid*;UNIX.gid*;unique*;
211 End.

For HTTP servers, the recognition of directories is even worse. Don't rely on it.


There are the generic copy algorithms Netfs.copy and Netfs.copy_into, which can also be used for HTTP and FTP.

For example, let's copy the file "/xyz" from fs1 to fs2, i.e. from an HTTP server to an FTP server:

Netfs.copy (fs1 :> Netfs.stream_fs) "/xyz" (fs2 : Netfs.stream_fs) "/xyz"

Iterating files

There is the generic file iterator Netfs.iter, which walks through the directory hierarchy on the server:

  ~pre:(fun name kind symkind -> ...)
  (fs2 :> Netfs.stream_fs)

Note that you may run into problems in conjunction with HTTP and FTP:

  • HTTP may redirect file accesses, and the iterator is not aware of this
  • FTP does not represent symlinks as such, and the iterator would follow symlink loops infinitely
  • Both HTTP and FTP have problems recognizing directories (see the remarks about globbing above).

Accessing HTTP via Http_client

The Http_client module is the real implementation of HTTP. It is asynchronous, which means it can do many tasks in parallel, but also needs special care when using it.

The tasks are organized as pipelines. This is actually an HTTP protocol feature - one can send the next request(s) to an HTTP server without having to wait for the response of the prior request. The pipeline is available as Ocaml class:

let p = new Http_client.pipeline

By adding requests, the pipeline is told to send them to the right server. If the server allows pipelining on HTTP level, this feature is exploited to speed up the accesses. Here, we submit two different GET requests:

let x1 = new Http_client.get ""
let x2 = new Http_client.get ""

p # add x1;
p # add x2

The objects x1 and x2 are instances of Http_client.http_call. They have lots of access methods for changing the request type and getting the returned response.

Now, after just adding the objects, nothing is done yet. You also have to start the pipeline:

p # run()

(Or, alternatively, do p#event_system, which is just the same.)

Now, you can get the fetched data with:

let d1 = x1 # response_body # value

Before looking at the value, you would normally check the status code of the response. There are a few possibilities:

How the pipeline works

The object p is actually several pipelines in one object. For each connected server, p keeps a small number of parallel connections (normally 2). Each connection is then driven in a pipelined way, if possible.

When running p, all the connections to the servers are created in parallel, and the communication is done in parallel.

Accessing FTP via Ftp_client

The Ftp_client is also an asynchronous implementation. It is a bit more difficult to exploit this, though, because Ftp_client is a bit simpler than Http_client.

Generally, a client can only connect to a single server, not to several at once. Also, there is no queue of pending requests - all submitted requests are immediately executed, and the next request can first be started when the previous has finished.

In synchrounous code, a file download looks like:

let client = new ftp_client()
let () = client # exec (connect_method ~host:"" ())
let () = client # exec 
	       ~get_password:(fun () -> "")
	       ~get_account:(fun () -> "")
let buffer = Buffer.create 1000 
let ch = new Netchannels.output_buffer buffer
let () = client # exec 
                ~file:(`NVFS "dir/baz")
                ~store:(fun _ -> `File_structure ch) 
let s = Buffer.contents buffer

As you see, this is just a sequence of exec calls. There is also an exec_e method allowing to start these operations as Uq_engines.engine, allowing asynchronous execution.

FTP has a number of subtle protocol options - like file transfer in several modes. Please refer to Ftp_client where these details are extensively documented.

You close the connection using the QUIT command with

let () = client # exec (quit_method())

or just run client # reset() to just shut the TCP connection down.

This web site is published by Informatikbüro Gerd Stolpmann
Powered by Caml