Plasma GitLab Archive
Projects Blog Knowledge


Netchannels Tutorial

Netchannels is one of the basic modules of this library, because it provides some very basic abstractions needed for many other functions of the library. The key abstractions Netchannels defines are the types in_obj_channel and out_obj_channel. Both are class types providing sequential access to byte streams, one for input, one for output. They are comparable to the types in_channel and out_channel of the standard library that allow access to files. However, there is one fundamental difference: in_channel and out_channel are restricted to resources that are available through file descriptors, whereas in_obj_channel and out_obj_channel are just class types, and by providing implementations for them any kind of resources can be accessed.


In some respect, Netchannels fixes a deficiency of the standard library. Look at the module Printf which defines six variants of the printf function:

val fprintf : out_channel -> ('a, out_channel, unit) format -> 'a
val printf : ('a, out_channel, unit) format -> 'a
val eprintf : ('a, out_channel, unit) format -> 'a
val sprintf : ('a, unit, string) format -> 'a
val bprintf : Buffer.t -> ('a, Buffer.t, unit) format -> 'a
val kprintf : (string -> string) -> ('a, unit, string) format -> 'a
It is possible to write into six different kinds of print targets. The basic problem of this style is that the provider of a service function like printf must define it for every commonly used print target. The other solution is that the provider defines only one version of the service function, but that the caller of the function arranges the polymorphism. A Netchannels-aware Printf would have only one variant of printf:
val printf : out_obj_channel -> ('a, out_obj_channel, unit) format -> 'a
The caller would create the right out_obj_channel object for the real print target:
let file_ch = new output_file (file : out_channel) in
printf file_ch ...
(printing into files), or:
let buffer_ch = new output_buffer (buf : Buffer.t) in
printf buffer_ch ...
(printing into buffers). Of course, this is only a hypothetical example. The point is that this library defines many parsers and printers, and that it is really a simplification for both the library and the user of the library to have this object encapsulation of I/O resources.

Programming with in_obj_channel

For example, let us program a function reading a data source line by line, and returning the sum of all lines which must be integer numbers. The argument ch is an open Netchannels.in_obj_channel, and the return value is the sum:

let sum_up (ch : in_obj_channel) =
  let sum = ref 0 in
    while true do
      let line = ch # input_line() in
      sum := !sum + int_of_string line
    assert false
    End_of_file ->
The interesting point is that the data source can be anything: a channel, a string, or any other class that implements the class type in_obj_channel.

This expression opens the file "data" and returns the sum of this file:

let ch = new input_channel (open_in "data") in
sum_up ch
The class Netchannels.input_channel is an implementation of the type in_obj_channel where every method of the class simply calls the corresponding function of the module Pervasives. (By the way, it would be a good idea to close the channel afterwards: ch#close_in(). We will discuss that below.)

This expression sums up the contents of a constant string:

let s = "1\n2\n3\n4" in
let ch = new input_string s in
sum_up ch
The class Netchannels.input_string is an implementation of the type in_obj_channel that reads from a string that is treated like a channel.

The effect of using the Netchannels module is that the same implementation sum_up can be used to read from multiple data sources, as it is sufficient to call the function with different implementations of in_obj_channel.

The details of in_obj_channel

The properties of any class that implements in_obj_channel can be summarized as follows:

  • After the object has been created (new), the netchannel is open. The netchannel remains open until it is explicitly closed (method close_in : unit -> unit). When you call a method of a closed netchannel, the exception Closed_channel is raised (even if you try to close the channel again).
  • The methods
      really_input : string -> int -> int -> unit
      input_char : unit -> char
      input_byte : unit -> int
      input_line : unit -> string
    work like their counterparts of the standard library. In particular, the end of file condition is signaled by rasising End_of_file.
  • The method
      input : string -> int -> int -> int
    works like its counterpart of the standard library, except that the end of the file is also signaled by End_of_file, and not by the return value 0.
  • The method pos_in : int returns the current byte position of the channel in a way that is logically consistent with the input methods: After reading n bytes, the method must return a position that is increased by n. Usually the position is zero after the object has been created, but this is not specified. Positions are available even for file descriptors that are not seekable.
  • There is intentionally no seek_in method. Seekable channels are currently out of scope, as netstring focuses on non-seekable channels.

Programming with out_obj_channel

The following function outputs the numbers of an int list sequentially on the passed netchannel:

let print_int_list (ch : out_obj_channel) l =
    (fun n ->
       ch # output_string (string_of_int n);
       ch # output_char '\n';
  ch # flush()
The following statements write the output into a file:
let ch = new output_channel (open_out "data") in
print_int_list ch [1;2;3]
And these statements write the output into a buffer:
let b = Buffer.create 16 in
let ch = new output_buffer b in
print_int_list ch [1;2;3]

Again, the caller of the function print_int_list determines the type of the output destination, and you do not need several functions for several types of destination.

The details of out_obj_channel

The properties of any class that implements out_obj_channel can be summarized as follows:

  • After the object has been created (new), the netchannel is open. The netchannel remains open until it is explicitly closed (method close_out : unit -> unit). When you call a method of a closed netchannel, the exception Closed_channel is raised (even if you try to close the channel again).
  • The methods
      output : string -> int -> int -> int
      really_output : string -> int -> int -> unit
      output_char : char -> unit
      output_byte : int -> unit
      output_string : string -> unit
    work like their counterparts of the standard library. There is usually an output buffer, but this is not specified. By calling flush : unit -> unit, the contents of the output buffer are forced to be written to the destination.
  • The method
      output_buffer : Buffer.t -> unit
    works like Buffer.output_channel, i.e. the contents of the buffer are printed to the channel.
  • The method
      output_channel : ?len:int -> in_obj_channel -> unit
    reads data from the argument in_obj_channel and prints them to the output channel. By default, the input channel is read until the EOF position. If the len argument is passed, at most this number of bytes are copied from the input channel to the output channel. The input channel remains open in all cases.
  • The method pos_out : int returns byte positions that are logically consistent: After writing n bytes, the method must return a position that is increased by n. Usually the position is zero after the object has been created, but this is not specified. Positions are available even for file descriptors that are not seekable.
  • There is intentionally no seek_out method. Seekable channels are currently out of scope, as netstring focuses on non-seekable channels.

How to close channels

As channels may use file descriptors for their implementation, it is very important that all open channels are closed after they have been used; otherwise the operating system will certainly get out of file descriptors. The simple way,

let ch = new <channel_class> args ... in
... do something ...
ch # close_in() or close_out()
is dangerous because an exception may be raised between channel creation and the close_* invocation. An elegant solution is to use with_in_obj_channel and with_out_obj_channel, as in:
with_in_obj_channel             (* or with_out_obj_channel *)
  (new <channel_class> ...)
  (fun ch ->
     ... do something ...
This programming idiom ensures that the channel is always closed after usage, even in the case of exceptions.

Complete examples:

let sum = with_in_obj_channel
            (new input_channel (open_in "data"))
            sum_up ;;

  (new output_channel (open_out "data"))
  (fun ch -> print_int_list ch ["1";"2";"3"]) ;;

Examples: HTML Parsing and Printing

In the Netstring library there are lots of parsers and printers that accept netchannels as data sources and destinations, respectively. One of them is the Nethtml module providing an HTML parser and printer. A few code snippets how to call them, just to get used to netchannels:

let html_document =
    (new input_channel (open_in "myfile.html"))
    Nethtml.parse ;;
  (new output_channel (open_out "otherfile.html"))
  (fun ch -> Nethtml.write ch html_document) ;;

Transactional Output Channels

Sometimes you do not want that generated output is directly sent to the underlying file descriptor, but rather buffered until you know that everything worked fine. Imagine you program a network service, and you want to return the result only when the computations are successful, and an error message otherwise. One way to achieve this effect is to manually program a buffer:

let network_service ch =
    let b = Buffer.create 16 in
    let ch' = new output_buffer b in
    ... computations, write results into ch' ...
    ch' # close_out;
    ch # output_buffer b
    error ->
      ... write error message to ch ...
There is a better way to do this, as there are transactional output channels. This type of netchannels provide a buffer for all written data like the above example, and only if data is explicitly committed it is copied to the real destination. Alternatively, you can also rollback the channel, i.e. delete the internal buffer. The signature of the type trans_out_obj_channel is:
class type trans_out_obj_channel = object
  inherit out_obj_channel
  method commit_work : unit -> unit
  method rollback_work : unit -> unit
They have the same methods as out_obj_channel plus commit_work and rollback_work. There are two implementations, one of them keeping the buffer in memory, and the other using a temporary file:
let ch' = new buffered_trans_channel ch
let ch' = new tempfile_trans_channel ch
In the latter case, there are optional arguments specifiying where the temporary file is created.

Now the network service would look like:

let network_service transaction_provider ch =
    let ch' = transaction_provider ch in
    ... computations, write results into ch' ...
    ch' # commit_work();
    ch' # close_out()     (* implies ch # close_out() *)
    error ->
      ch' # rollback_work();
      ... write error message to ch' ...
      ch' # commit_work();
      ch' # close_out()   (* implies ch # close_out() *)
You can program this function without specifying which of the two implementations is used. Just call this function as
network_service (new buffered_trans_channel) ch
network_service (new tempfile_trans_channel) ch
to determine the type of transaction buffer.

Some details:

  • The method commit_work copies all uncommitted data to the underlying channel, and flushes all buffers.
  • When rollback_work is called the uncommitted data are deleted.
  • The method flush does not have any effect.
  • The reported position adds the committed and the uncommitted amounts of data. This means that rollback_work resets the position to the value of the last commit_work call.
  • When the transactional channel is closed, the underlying channel is closed, too. By default, the uncommitted data is deleted, but the current implementations can optionally commit data in this case.

Pipes and Filters

The class pipe is an in_obj_channel and an out_obj_channel at the same time (i.e. the class has the type io_obj_channel). A pipe has two endpoints, one for reading and one for writing (similar in concept to the pipes provided by the operating system, but note that our pipes have nothing to do with the OS pipes). Of course, you cannot read and write at the same time, so there must be an internal buffer storing the data that have been written but not yet read. How can such a construction be useful? Imagine you have two routines that run alternately, and one is capable of writing into netchannels, and the other can read from a netchannel. Pipes are the missing communication link in this situation, because the writer routine can output into the pipe, and the reader routine can read from the buffer of the pipe. In the following example, the writer outputs numbers from 1 to 100, and the reader sums them up:

let pipe = new pipe() ;;
let k = ref 1 ;;
let writer() =
  if !k <= 100 then (
    pipe # output_string (string_of_int !k);
    incr k;
    if !k > 100 then pipe # close_out() else pipe # output_char '\n';
  ) ;;
let sum = ref 0 ;;
let reader() =
  let line = pipe # input_line() in
  sum := !sum + int_of_string line ;;
  while true do
  End_of_file ->
    () ;;
The writer function prints the numbers into the pipe, and the reader function reads them in. By closing only the output end Of the pipe the writer signals the end of the stream, and the input_line method raises the exception End_of_file.

Of course, this example is very simple. What does happen when more is printed into the pipe than read? The internal buffer grows. What does happen when more is tried to read from the pipe than available? The input methods signal this by raising the special exception Buffer_underrun. Unfortunately, handling this exception can be very complicated, as the reader must be able to deal with partial reads.

This could be solved by using the Netstream module. A netstream is another extension of in_obj_channel that allows one to look ahead, i.e. you can look at the bytes that will be read next, and use this information to decide whether enough data are available or not. Netstreams are explained in another chapter of this manual.

Pipes have another feature that makes them useful even for "normal" programming. You can specify a conversion function that is called when data is to be transferred from the writing end to the reading end of the pipe. The module Netencoding.Base64 defines such a pipe that converts data: The class encoding_pipe automatically encodes all bytes written into it by the Base64 scheme:

let pipe = new Netencoding.Base64.encoding_pipe() ;;
pipe # output_string "Hello World";
pipe # close_out() ;;
let s = pipe # input_line() ;;
s has now the value "SGVsbG8gV29ybGQ=", the encoded form of the input. This kind of pipe has the same interface as the basic pipe class, and the same problems to use it. Fortunately, the Netstring library has another facility simplifying the usage of pipes, namely filters.

There are two kinds of filters: The class Netchannels.output_filter redirects data written to an out_obj_channel through a pipe, and the class Netchannels.input_filter arranges that data read from an in_obj_channel flows through a pipe. An example makes that clearer. Imagine you have a function write_results that writes the results of a computation into an out_obj_channel. Normally, this channel is simply a file:

  (new output_channel (open_out "results"))
Now you want that the file is Base64-encoded. This can be arranged by calling write_results differently:
let pipe = new Netencoding.Base64.encoding_pipe() in
  (new output_channel (open_out "results"))
  (fun ch ->
    let ch' = new output_filter pipe ch in
    write_results ch';
    ch' # close_out()
Now any invocation of an output method for ch' actually prints into the filter, which redirects the data through the pipe, thus encoding them, and finally passing the encoded data to the underlying channel ch. Note that you must close ch' to ensure that all data are filtered, it is not sufficient to flush output.

It is important to understand why filters must be closed to work properly. The problem is that the Base64 encoding converts triples of three bytes into quadruples of four bytes. Because not every string to convert is a multiple of three, there are special rules how to handle the exceeding one or two bytes at the end. The pipe must know the end of the input data in order to apply these rules correctly. If you only flush the filter, the exceeding bytes would simply remain in the internal buffer, because it is possible that more bytes follow. By closing the filter, you indicate that the definite end is reached, and the special rules for trailing data must be performed. \- Many conversions have similar problems, and because of this it is a good advice to always close output filters after usage.

There is not only the class output_filter but also input_filter. This class can be used to perform conversions while reading from a file. Note that you often do not need to close input filters, because input channels can signal the end by raising End_of_file, so the mentioned problems usually do not occur.

There are a number of predefined conversion pipes:

  • Netencoding.Base64.encoding_pipe: Performs Base64 encoding
  • Netencoding.Base64.decoding_pipe: Performs Base64 decoding
  • Netencoding.QuotedPrintable.encoding_pipe: Performs QuotedPrintable encoding
  • Netencoding.QuotedPrintable.decoding_pipe: Performs QuotedPrintable decoding
  • Netconversion.conversion_pipe: Converts the character encoding form charset A to charset B

Defining Classes for Object Channels

As subtyping and inheritance are orthogonal in O'Caml, you can simply create your own netchannels by defining classes that match the in_obj_channel or out_obj_channel types. E.g.

class my_in_channel : in_obj_channel =
object (self)
  method input s pos len = ...
  method close_in() = ...
  method pos_in = ...
  method really_input s pos len = ...
  method input_char() = ...
  method input_line() = ...
  method input_byte() = ...

Of course, this is non-trivial, especially for the in_obj_channel case. Fortunately, the Netchannels module includes a "construction kit" that allows one to define a channel class from only a few methods. A closer look at in_obj_channel and out_obj_channel shows that some methods can be derived from more fundamental methods. The following class types include only the fundamental methods:

class type raw_in_channel = object
  method input : string -> int -> int -> int
  method close_in : unit -> unit
  method pos_in : int
class type raw_out_channel = object
  method output : string -> int -> int -> int
  method close_out : unit -> unit
  method pos_out : int
  method flush : unit -> unit

In order to define a new class, it is sufficient to define this raw version of the class, and to lift it to the full functionality. For example, to define my_in_channel:

class my_raw_in_channel : raw_in_channel =
object (self)
  method input s pos len = ...
  method close_in() = ...
  method pos_in = ...
class my_in_channel =
  in_obj_channel_delegation (lift_in (`Raw(new my_raw_in_channel)))

The function Netchannels.lift_in can lift several forms of incomplete channel objects to the full class type in_obj_channel. There is also the corresponding function Netchannels.lift_out. Note that lifting adds by default another internal buffer to the channel that must be explicitly turned off when it is not wanted. The rationale for this buffer is that it avoids some cases with extremely poor performance which might be surprising for many users.

The class in_obj_channel_delegation is just an auxiliary construction to turn the in_obj_channel object returned by lift_in again into a class.

Some FAQ

  • Netchannels add further layers on top of the built-in channels or file descriptors. Does this make them slow?

    Of course, Netchannels are slower than the underlying built-in I/O facilities. There is at least one, but often even more than one method call until the data is transferred to or from the final I/O target. This costs time, and it is a good idea to reduce the number of method calls for maximum speed. Especially the character- or byte-based method calls should be avoided, it is better to collect data and pass them in larger chunks. This reduces the number of method calls that are needed to transfer a block of data.

    However, some classes implement buffers themselves, and data are only transferred when the buffers are full (or empty). The overhead for the extra method calls is small for these classes. The classes that implement their own buffers are the transactional channels, the pipes, and all the classes with "buffer" in their name.

    Netchannels are often stacked, i.e. one netchannel object transfers data to an underlying object, and this object passes the data to further objects. Often buffers are involved, and data are copied between buffers several times. Of course, these copies can reduce the speed, too.

  • Why do Netchannels not support seeking?

    Netchannels were invented to support the implementation of network protocols. Network endpoints are not seekable.

  • What about printf and scanf?

    In principle, methods for printf and scanf could be added to out_obj_channel and in_obj_channel, respectively, as recent versions of O'Caml added the necessary language means (polymorphic methods, kprintf, kscanf). However, polymorphic methods work only well when the type of the channel object is always annotated (e.g. as (ch : out_obj_channel) # printf ...), so this is not that much better than ch # output_string (sprintf ...).

  • Can I pass an in_obj_channel to an ocamllex-generated lexer?

    Yes, just call Netchannels.lexbuf_of_in_obj_channel to turn the in_obj_channel into a lexbuf.

  • Do Netchannels support non-blocking I/O?

    Yes and no. Yes, because you can open a descriptor in non-blocking mode, and create a netchannel from it. When the program would block, the input and output methods return 0 to indicate this. However, the non-raw methods cannot cope with these situations.

  • Do Netchannels support multiplexed I/O?

    No, there is no equivalent to on the level of netchannels.

  • Can I use Netchannels in multi-threaded programs?

    Yes. However, shared netchannels are not locked, and strange things can happen when netchannels are used by several threads at the same time.

  • Can I use pipes to communicate between threads?

    This could be made work, but it is currently not the case. A multithreading-aware wrapper around pipes could do the job.

  • Pipes call external programs to do their job, don't they?

    No, they do not call external programs, nor do they need any parallel execution threads. Pipes are just a tricky way of organizing buffers.

  • How do I define my own conversion pipe?

    Look at the sources, it includes several examples of conversion pipes.

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