Warning: Reason support is experimental. We are looking for beta-tester and contributors.

Lwt manual

Introduction

When writing a program, a common developer's task is to handle I/O operations. Indeed, most software interacts with several different resources, such as:

  • the kernel, by doing system calls,
  • the user, by reading the keyboard, the mouse, or any input device,
  • a graphical server, to build graphical user interface,
  • other computers, by using the network,
  • ...and so on.

When this list contains only one item, it is pretty easy to handle. However as this list grows it becomes harder and harder to make everything work together. Several choices have been proposed to solve this problem:

  • using a main loop, and integrating all components we are interacting with into this main loop.
  • using preemptive system threads

Both solutions have their advantages and their drawbacks. For the first one, it may work, but it becomes very complicated to write a piece of asynchronous sequential code. The typical example is graphical user interfaces freezing and not redrawing themselves because they are waiting for some blocking part of the code to complete.

If you already wrote code using preemptive threads, you should know that doing it right with threads is a difficult job. Moreover, system threads consume non-negligible resources, and so you can only launch a limited number of threads at the same time. Thus, this is not a general solution.

Lwt offers a third alternative. It provides promises, which are very fast: a promise is just a reference that will be filled asynchronously, and calling a function that returns a promise does not require a new stack, new process, or anything else. It is just a normal, fast, function call. Promises compose nicely, allowing us to write highly asynchronous programs.

In the first part, we will explain the concepts of Lwt, then we will describe the main modules Lwt consists of.

The Lwt core library

In this section we describe the basics of Lwt. It is advised to start utop and try the given code examples.

Lwt concepts

Let's take a classic function of the Pervasives module:

# Pervasives.input_char;;
- : in_channel -> char = <fun>

This function will wait for a character to come on the given input channel, and then return it. The problem with this function is that it is blocking: while it is being executed, the whole program will be blocked, and other events will not be handled until it returns.

Now, let's look at the lwt equivalent:

# Lwt_io.read_char;;
- : Lwt_io.input_channel -> char Lwt.t = <fun>

As you can see, it does not return just a character, but something of type char Lwt.t. The type 'a Lwt.t is the type of promises that can be fulfilled later with a value of type 'a. Lwt_io.read_char will try to read a character from the given input channel and immediately return a promise, without blocking, whether a character is available or not. If a character is not available, the promise will just not be fulfilled yet.

Now, let's see what we can do with a Lwt promise. The following code creates a pipe, creates a promise that is fulfilled with the result of reading the input side:

# let ic, oc = Lwt_io.pipe ();;
val ic : Lwt_io.input_channel = <abstr>
val oc : Lwt_io.output_channel = <abstr>
# let p = Lwt_io.read_char ic;;
val p : char Lwt.t = <abstr>

We can now look at the state of our newly created promise:

# Lwt.state p;;
- : char Lwt.state = Lwt.Sleep

A promise may be in one of the following states:

  • Return x, which means that the promise has been fulfilled with the value x. This usually implies that the asynchronous operation, that you started by calling the function that returned the promise, has completed successfully.
  • Fail exn, which means that the promise has been rejected with the exception exn. This usually means that the asynchronous operation associated with the promise has failed.
  • Sleep, which means that the promise is has not yet been fulfilled or rejected, so it is pending.

The above promise p is pending because there is nothing yet to read from the pipe. Let's write something:

# Lwt_io.write_char oc 'a';;
- : unit Lwt.t = <abstr>
# Lwt.state p;;
- : char Lwt.state = Lwt.Return 'a'

So, after we write something, the reading promise has been fulfilled with the value 'a'.

Primitives for promise creation

There are several primitives for creating Lwt promises. These functions are located in the module Lwt.

Here are the main primitives:

  • Lwt.return : 'a -> 'a Lwt.t
    creates a promise which is already fulfilled with the given value
  • Lwt.fail : exn -> 'a Lwt.t
    creates a promise which is already rejected with the given exception
  • Lwt.wait : unit -> 'a Lwt.t * 'a Lwt.u
    creates a pending promise, and returns it, paired with a resolver (of type 'a Lwt.u), which must be used to resolve (fulfill or reject) the promise.

To resolve a pending promise, use one of the following functions:

  • Lwt.wakeup : 'a Lwt.u -> 'a -> unit
    fulfills the promise with a value.
  • Lwt.wakeup_exn : 'a Lwt.u -> exn -> unit
    rejects the promise with an exception.

Note that it is an error to try to resolve the same promise twice. Lwt will raise Invalid_argument if you try to do so.

With this information, try to guess the result of each of the following expressions:

# Lwt.state (Lwt.return 42);;
# Lwt.state (Lwt.fail Exit);;
# let p, r = Lwt.wait ();;
# Lwt.state p;;
# Lwt.wakeup r 42;;
# Lwt.state p;;
# let p, r = Lwt.wait ();;
# Lwt.state p;;
# Lwt.wakeup_exn r Exit;;
# Lwt.state p;;

Primitives for promise composition

The most important operation you need to know is bind:

val bind : 'a Lwt.t -> ('a -> 'b Lwt.t) -> 'b Lwt.t

bind p f creates a promise which waits for p to become become fulfilled, then passes the resulting value to f. If p is a pending promise, then bind p f will be a pending promise too, until p is resolved. If p is rejected, then the resulting promise will be rejected with the same exception. For example, consider the following expression:

Lwt.bind
  (Lwt_io.read_line Lwt_io.stdin)
  (fun str -> Lwt_io.printlf "You typed %S" str)

This code will first wait for the user to enter a line of text, then print a message on the standard output.

Similarly to bind, there is a function to handle the case when p is rejected:

val catch : (unit -> 'a Lwt.t) -> (exn -> 'a Lwt.t) -> 'a Lwt.t

catch f g will call f (), then wait for it to become resolved, and if it was rejected with an exception exn, call g exn to handle it. Note that both exceptions raised with Pervasives.raise and Lwt.fail are caught by catch.

Cancelable promises

In some case, we may want to cancel a promise. For example, because it has not resolved after a timeout. This can be done with cancelable promises. To create a cancelable promise, you must use the Lwt.task function:

val task : unit -> 'a Lwt.t * 'a Lwt.u

It has the same semantics as Lwt.wait, except that the pending promise can be canceled with Lwt.cancel:

val cancel : 'a Lwt.t -> unit

The promise will then be rejected with the exception Lwt.Canceled. To execute a function when the promise is canceled, you must use Lwt.on_cancel:

val on_cancel : 'a Lwt.t -> (unit -> unit) -> unit

Note that canceling a promise does not automatically cancel the asynchronous operation that is going to resolve it. It does, however, prevent any further chained operations from running. The asynchronous operation associated with a promise can only be canceled if its implementation has taken care to set an on_cancel callback on the promise that it returned to you. In practice, most operations (such as system calls) can't be canceled once they are started anyway, so promise cancellation is useful mainly for interrupting future operations once you know that a chain of asynchronous operations will not be needed.

It is also possible to cancel a promise which has not been created directly by you with Lwt.task. In this case, the deepest cancelable promise that the given promise depends on will be canceled.

For example, consider the following code:

# let p, r = Lwt.task ();;
val p : '_a Lwt.t = <abstr>
val r : '_a Lwt.u = <abstr>
# let p' = Lwt.bind p (fun x -> Lwt.return (x + 1));;
val p' : int Lwt.t = <abstr>

Here, cancelling p' will in fact cancel p, rejecting it with Lwt.Canceled. Lwt.bind will then propagate the exception forward to p':

# Lwt.cancel p';;
- : unit = ()
# Lwt.state p;;
- : int Lwt.state = Lwt.Fail Lwt.Canceled
# Lwt.state p';;
- : int Lwt.state = Lwt.Fail Lwt.Canceled

It is possible to prevent a promise from being canceled by using the function Lwt.protected:

val protected : 'a Lwt.t -> 'a Lwt.t

Canceling (protected p) will have no effect on p.

Primitives for concurrent composition

We now show how to compose several promises concurrently. The main functions for this are in the Lwt module: join, choose and pick.

The first one, join takes a list of promises and returns a promise that is waiting for all of them to resolve:

val join : unit Lwt.t list -> unit Lwt.t

Moreover, if at least one promise is rejected, join l will be rejected with the same exception as the first one, after all the promises are resolved.

Conversely, choose waits for at least one promise to become resolved, then resolves with the same value or exception:

val choose : 'a Lwt.t list -> 'a Lwt.t

For example:

# let p1, r1 = Lwt.wait ();;
val p1 : '_a Lwt.t = <abstr>
val r1 : '_a Lwt.u = <abstr>
# let p2, r2 = Lwt.wait ();;
val p2 : '_a Lwt.t = <abstr>
val r2 : '_a Lwt.u = <abstr>
# let p3 = Lwt.choose [p1; p2];;
val p3 : '_a Lwt.t = <abstr>
# Lwt.state p3;;
- : '_a Lwt.state = Lwt.Sleep
# Lwt.wakeup r2 42;;
- : unit = ()
# Lwt.state p3;;
- : int Lwt.state = Lwt.Return 42

The last one, pick, is the same as choose, except that it tries to cancel all other promises when one resolves. Promises created via Lwt.wait() are not cancellable and are thus not cancelled.

Rules

A callback, like the f that you might pass to Lwt.bind, is an ordinary OCaml function. Lwt just handles ordering calls to these functions.

Lwt uses some preemptive threading internally, but all of your code runs in the main thread, except when you explicitly opt into additional threads with Lwt_preemptive.

This simplifies reasoning about critical sections: all the code in one callback cannot be interrupted by any of the code in another callback. However, it also carries the danger that if a single callback takes a very long time, it will not give Lwt a chance to run your other callbacks. In particular:

  • do not write functions that may take time to complete, without splitting them up using Lwt.pause or performing some Lwt I/O,
  • do not do I/O that may block, otherwise the whole program will hang inside that callback. You must instead use the asynchronous I/O operations provided by Lwt.

The syntax extension

Lwt offers a PPX syntax extension which increases code readability and makes coding using Lwt easier. The syntax extension is documented here.

To use the PPX syntax extension, add the lwt_ppx package when compiling:

$ ocamlfind ocamlc -package lwt_ppx -linkpkg -o foo foo.ml

Or, in utop:

# #require "lwt_ppx";;

lwt_ppx is distributed in a separate opam package of that same name.

For a brief overview of the syntax, see the Correspondence table below.

Correspondence table

without Lwt with Lwt
let pattern1 = expr1 let%lwt pattern1 = expr1
and pattern2 = expr2 and pattern2 = expr2
... ...
and patternn = exprn in and patternn = exprn in
expr expr
try try%lwt
expr expr
with with
| pattern1 -> expr1 | pattern1 -> expr1
| pattern2 -> expr2 | pattern2 -> expr2
... ...
| patternn -> exprn | patternn -> exprn
for ident = exprinit to exprfinal do for%lwt ident = exprinit to exprfinal do
expr expr
done done
while expr do while%lwt expr do
expr expr
done done
assert expr assert%lwt expr
match expr with match%lwt expr with
| pattern1 -> expr1 | pattern1 -> expr1
| pattern2 -> expr2 | pattern2 -> expr2
... ...
| patternn -> exprn | patternn -> exprn
if expr then expr else expr if%lwt expr then expr else expr
raise exn [%lwt raise exn]

Backtrace support

If an exception is raised inside a callback called by Lwt, the backtrace provided by OCaml will not be very useful. It will end inside the Lwt scheduler instead of continuing into the code that started the operations that led to the callback call. To avoid this, and get good backtraces from Lwt, use the syntax extension. The let%lwt construct will properly propagate backtraces.

As always, to get backtraces from an OCaml program, you need to either declare the environment variable OCAMLRUNPARAM=b or call Printexc.record_backtrace true at the start of your program, and be sure to compile it with -g. Most modern build systems add -g by default.

Other modules of the core library

The core library contains several modules that only depend on Lwt. The following naming convention is used in Lwt: when a function takes as argument a function, returning a promise, that is going to be executed sequentially, it is suffixed with ``_s''. And when it is going to be executed concurrently, it is suffixed with ``_p''. For example, in the Lwt_list module we have:

val map_s : ('a -> 'b Lwt.t) -> 'a list -> 'b list Lwt.t
val map_p : ('a -> 'b Lwt.t) -> 'a list -> 'b list Lwt.t

Mutexes

Lwt_mutex provides mutexes for Lwt. Its use is almost the same as the Mutex module of the thread library shipped with OCaml. In general, programs using Lwt do not need a lot of mutexes, because callbacks run without preempting each other. They are only useful for synchronising or sequencing complex operations spread over multiple callback calls.

Lists

The Lwt_list module defines iteration and scanning functions over lists, similar to the ones of the List module, but using functions that return a promise. For example:

val iter_s : ('a -> unit Lwt.t) -> 'a list -> unit Lwt.t
val iter_p : ('a -> unit Lwt.t) -> 'a list -> unit Lwt.t

In iter_s f l, iter_s will call f on each elements of l, waiting for resolution between each element. On the contrary, in iter_p f l, iter_p will call f on all elements of l, only then wait for all the promises to resolve.

Data streams

Lwt streams are used in a lot of places in Lwt and its submodules. They offer a high-level interface to manipulate data flows.

A stream is an object which returns elements sequentially and lazily. Lazily means that the source of the stream is touched only for new elements when needed. This module contains a lot of stream transformation, iteration, and scanning functions.

The common way of creating a stream is by using Lwt_stream.from or by using Lwt_stream.create:

val from : (unit -> 'a option Lwt.t) -> 'a Lwt_stream.t
val create : unit -> 'a Lwt_stream.t * ('a option -> unit)

As for streams of the standard library, from takes as argument a function which is used to create new elements.

create returns a function used to push new elements into the stream and the stream which will receive them.

For example:

# let stream, push = Lwt_stream.create ();;
val stream : '_a Lwt_stream.t = <abstr>
val push : '_a option -> unit = <fun>
# push (Some 1);;
- : unit = ()
# push (Some 2);;
- : unit = ()
# push (Some 3);;
- : unit = ()
# Lwt.state (Lwt_stream.next stream);;
- : int Lwt.state = Lwt.Return 1
# Lwt.state (Lwt_stream.next stream);;
- : int Lwt.state = Lwt.Return 2
# Lwt.state (Lwt_stream.next stream);;
- : int Lwt.state = Lwt.Return 3
# Lwt.state (Lwt_stream.next stream);;
- : int Lwt.state = Lwt.Sleep

Note that streams are consumable. Once you take an element from a stream, it is removed from the stream. So, if you want to iterate two times over a stream, you may consider ``cloning'' it, with Lwt_stream.clone. Cloned stream will return the same elements in the same order. Consuming one will not consume the other. For example:

# let s = Lwt_stream.of_list [1; 2];;
val s : int Lwt_stream.t = <abstr>
# let s' = Lwt_stream.clone s;;
val s' : int Lwt_stream.t = <abstr>
# Lwt.state (Lwt_stream.next s);;
- : int Lwt.state = Lwt.Return 1
# Lwt.state (Lwt_stream.next s);;
- : int Lwt.state = Lwt.Return 2
# Lwt.state (Lwt_stream.next s');;
- : int Lwt.state = Lwt.Return 1
# Lwt.state (Lwt_stream.next s');;
- : int Lwt.state = Lwt.Return 2

Mailbox variables

The Lwt_mvar module provides mailbox variables. A mailbox variable, also called a ``mvar'', is a cell which may contain 0 or 1 element. If it contains no elements, we say that the mvar is empty, if it contains one, we say that it is full. Adding an element to a full mvar will block until one is taken. Taking an element from an empty mvar will block until one is added.

Mailbox variables are commonly used to pass messages between chains of callbacks being executed concurrently.

Note that a mailbox variable can be seen as a pushable stream with a limited memory.

Running an Lwt program

An Lwt computation you have created will give you something of type Lwt.t, a promise. However, even though you have the promise, the computation may not have run yet, and the promise might still be pending.

For example if your program is just:

let _ = Lwt_io.printl "Hello, world!"

you have no guarantee that the promise for writing "Hello, world!" on the terminal will be resolved before the program exits. In order to wait for the promise to resolve, you have to call the function Lwt_main.run:

val Lwt_main.run : 'a Lwt.t -> 'a

This function waits for the given promise to resolve and returns its result. In fact it does more than that; it also runs the scheduler which is responsible for making asynchronous computations progress when events are received from the outside world.

So basically, when you write a Lwt program, you must call Lwt_main.run on your top-level, outer-most promise. For instance:

let () = Lwt_main.run (Lwt_io.printl "Hello, world!")

Note that you must not make nested calls to Lwt_main.run. It cannot be used anywhere else to get the result of a promise.

The lwt.unix library

The package lwt.unix contains all Unix-dependent modules of Lwt. Among all its features, it implements Lwt-friendly, non-blocking versions of functions of the OCaml standard and Unix libraries.

Unix primitives

Module Lwt_unix provides non-blocking system calls. For example, the Lwt counterpart of Unix.read is:

val read : file_descr -> string -> int -> int -> int Lwt.t

Lwt_io provides features similar to buffered channels of the standard library (of type in_channel or out_channel), but with non-blocking semantics.

Lwt_gc allows you to register a finalizer that returns a promise. At the end of the program, Lwt will wait for all these finalizers to resolve.

The Lwt scheduler

Operations doing I/O have to be resumed when some events are received by the process, so they can resolve their associated pending promises. For example, when you read from a file descriptor, you may have to wait for the file descriptor to become readable if no data are immediately available on it.

Lwt contains a scheduler which is responsible for managing multiple operations waiting for events, and restarting them when needed. This scheduler is implemented by the two modules Lwt_engine and Lwt_main. Lwt_engine is a low-level module, it provides a signature for custom I/O multiplexers as well as two built-in implementations, libev and select. The signature is given by the class Lwt_engine.t.

libev is used by default on Linux, because it supports any number of file descriptors, while select supports only 1024. libev also much more efficient. On Windows, Unix.select is used because libev does not work properly. The user may change the backend in use at any time.

If you see an Invalid_argument error on Unix.select, it may be because the 1024 file descriptor limit was exceeded. Try switching to libev, if possible.

The engine can also be used directly in order to integrate other libraries with Lwt. For example, GTK needs to be notified when some events are received. If you use Lwt with GTK you need to use the Lwt scheduler to monitor GTK sources. This is what is done by the Lwt_glib library.

The Lwt_main module contains the main loop of Lwt. It is run by calling the function Lwt_main.run:

val Lwt_main.run : 'a Lwt.t -> 'a

This function continuously runs the scheduler until the promise passed as argument is resolved.

To make sure Lwt is compiled with libev support, tell opam that the library is available on the system by installing the conf-libev package. You may get the actual library with your system package manager:

  • brew install libev on MacOSX,
  • apt-get install libev-dev on Debian/Ubuntu, or
  • yum install libev-devel on CentOS, which requires to set export C_INCLUDE_PATH=/usr/include/libev/ and export LIBRARY_PATH=/usr/lib64/ before calling opam install conf-libev.

Logging

For logging, we recommend the logs package from opam, which includes an Lwt-aware module Logs_lwt.

The Lwt.react library

The Lwt_react module provides helpers for using the react library with Lwt. It extends the React module by adding Lwt-specific functions. It can be used as a replacement of React. For example you can add at the beginning of your program:

open Lwt_react

instead of:

open React

or:

module React = Lwt_react

Among the added functionalities we have Lwt_react.E.next, which takes an event and returns a promise which will be pending until the next occurrence of this event. For example:

# open Lwt_react;;
# let event, push = E.create ();;
val event : '_a React.event = <abstr>
val push : '_a -> unit = <fun>
# let p = E.next event;;
val p : '_a Lwt.t = <abstr>
# Lwt.state p;;
- : '_a Lwt.state = Lwt.Sleep
# push 42;;
- : unit = ()
# Lwt.state p;;
- : int Lwt.state = Lwt.Return 42

Another interesting feature is the ability to limit events (resp. signals) from occurring (resp. changing) too often. For example, suppose you are doing a program which displays something on the screen each time a signal changes. If at some point the signal changes 1000 times per second, you probably don't want to render it 1000 times per second. For that you use Lwt_react.S.limit:

val limit : (unit -> unit Lwt.t) -> 'a React.signal -> 'a React.signal

Lwt_react.S.limit f signal returns a signal which varies as signal except that two consecutive updates are separated by a call to f. For example if f returns a promise which is pending for 0.1 seconds, then there will be no more than 10 changes per second:

open Lwt_react

let draw x =
  (* Draw the screen *)
  ...

let () =
  (* The signal we are interested in: *)
  let signal = ... in

  (* The limited signal: *)
  let signal' = S.limit (fun () -> Lwt_unix.sleep 0.1) signal in

  (* Redraw the screen each time the limited signal change: *)
  S.notify_p draw signal'

Other libraries

Detaching computation to preemptive threads

It may happen that you want to run a function which will take time to compute or that you want to use a blocking function that cannot be used in a non-blocking way. For these situations, Lwt allows you to detach the computation to a preemptive thread.

This is done by the module Lwt_preemptive of the lwt.unix package which maintains a pool of system threads. The main function is:

val detach : ('a -> 'b) -> 'a -> 'b Lwt.t

detach f x will execute f x in another thread and return a pending promise, usable from the main thread, which will be fulfilled with the result of the preemptive thread.

If you want to trigger some Lwt operations from your detached thread, you have to call back into the main thread using Lwt_preemptive.run_in_main:

val run_in_main : (unit -> 'a Lwt.t) -> 'a

This is roughly the equivalent of Lwt.main_run, but for detached threads, rather than for the whole process. Note that you must not call Lwt_main.run in a detached thread.

SSL support

The library Lwt_ssl allows use of SSL asynchronously.

Writing stubs using Lwt

Thread-safe notifications

If you want to notify the main thread from another thread, you can use the Lwt thread safe notification system. First you need to create a notification identifier (which is just an integer) from the OCaml side using the Lwt_unix.make_notification function, then you can send it from either the OCaml code with Lwt_unix.send_notification function, or from the C code using the function lwt_unix_send_notification (defined in lwt_unix_.h).

Notifications are received and processed asynchronously by the main thread.

Jobs

For operations that cannot be executed asynchronously, Lwt uses a system of jobs that can be executed in a different threads. A job is composed of three functions:

  • A stub function to create the job. It must allocate a new job structure and fill its [worker] and [result] fields. This function is executed in the main thread. The return type for the OCaml external must be of the form 'a job.
  • A function which executes the job. This one may be executed asynchronously in another thread. This function must not:
    • access or allocate OCaml block values (tuples, strings, ...),
    • call OCaml code.
  • A function which reads the result of the job, frees resources and returns the result as an OCaml value. This function is executed in the main thread.

With Lwt < 2.3.3, 4 functions (including 3 stubs) were required. It is still possible to use this mode but it is deprecated.

We show as example the implementation of Lwt_unix.mkdir. On the C side we have:

/**/
/* Structure holding informations for calling [mkdir]. */
struct job_mkdir {
  /* Informations used by lwt.
     It must be the first field of the structure. */
  struct lwt_unix_job job;
  /* This field store the result of the call. */
  int result;
  /* This field store the value of [errno] after the call. */
  int errno_copy;
  /* Pointer to a copy of the path parameter. */
  char* path;
  /* Copy of the mode parameter. */
  int mode;
  /* Buffer for storing the path. */
  char data[];
};

/* The function calling [mkdir]. */
static void worker_mkdir(struct job_mkdir* job)
{
  /* Perform the blocking call. */
  job->result = mkdir(job->path, job->mode);
  /* Save the value of errno. */
  job->errno_copy = errno;
}

/* The function building the caml result. */
static value result_mkdir(struct job_mkdir* job)
{
  /* Check for errors. */
  if (job->result < 0) {
    /* Save the value of errno so we can use it
       once the job has been freed. */
    int error = job->errno_copy;
    /* Copy the contents of job->path into a caml string. */
    value string_argument = caml_copy_string(job->path);
    /* Free the job structure. */
    lwt_unix_free_job(&job->job);
    /* Raise the error. */
    unix_error(error, "mkdir", string_argument);
  }
  /* Free the job structure. */
  lwt_unix_free_job(&job->job);
  /* Return the result. */
  return Val_unit;
}

/* The stub creating the job structure. */
CAMLprim value lwt_unix_mkdir_job(value path, value mode)
{
  /* Get the length of the path parameter. */
  mlsize_t len_path = caml_string_length(path) + 1;
  /* Allocate a new job. */
  struct job_mkdir* job =
    (struct job_mkdir*)lwt_unix_new_plus(struct job_mkdir, len_path);
  /* Set the offset of the path parameter inside the job structure. */
  job->path = job->data;
  /* Copy the path parameter inside the job structure. */
  memcpy(job->path, String_val(path), len_path);
  /* Initialize function fields. */
  job->job.worker = (lwt_unix_job_worker)worker_mkdir;
  job->job.result = (lwt_unix_job_result)result_mkdir;
  /* Copy the mode parameter. */
  job->mode = Int_val(mode);
  /* Wrap the structure into a caml value. */
  return lwt_unix_alloc_job(&job->job);
}

and on the ocaml side:

(* The stub for creating the job. *)
external mkdir_job : string -> int -> unit job = "lwt_unix_mkdir_job"

(* The ocaml function. *)
let mkdir name perms = Lwt_unix.run_job (mkdir_job name perms)