Module `Async_kernel__.Mvar`

An Mvar is a mutable location that is either empty or contains a value. One can put or set the value, and wait on value_available for the location to be filled in either way.

Having an Mvar.Read_write.t gives the capability to mutate the mvar.

The key difference between an Mvar and an Ivar is that an Mvar may be filled multiple times.

This implementation of Mvar also allows one to replace the value without any guarantee that the reading side has seen it. This is useful in situations where last-value semantics are desired (e.g., you want to signal whenever a config file is updated, but only care about the most recent contents).

An Mvar can also be used as a baton-passing mechanism between a producer and consumer. For instance, a producer reading from a socket and producing a set of deserialized messages can put the batch from a single read into an Mvar and can wait for taken to return as a pushback mechanism. The consumer meanwhile waits on value_available. This way the natural batch size is passed between the two sub-systems with minimal overhead.

type ('a, -'phantom) t

val sexp_of_t : ('a -> Ppx_sexp_conv_lib.Sexp.t) -> ('phantom -> Ppx_sexp_conv_lib.Sexp.t) -> ('a, 'phantom) t -> Ppx_sexp_conv_lib.Sexp.t

module Read_write : sig ... end

module Read_only : sig ... end

val create : unit -> 'a Read_write.t
val is_empty : (_, _) t -> bool
val put : ('a, [> Core_kernel.write ]) t -> 'a -> unit Async_kernel.Deferred.t: put t a waits until is_empty t, and then does set t a. If there are multiple concurrent puts, there is no fairness guarantee (i.e., puts may happen out of order or may be starved).

val set : ('a, [> Core_kernel.write ]) t -> 'a -> unit: set t a sets the value in t to a, even if not (is_empty t). This is useful if you want takers to have last-value semantics.

val update : ('a, Core_kernel.read_write) t -> f:('a option -> 'a) -> unit: update t ~f applies f to the value in t and sets t to the result. This is useful if you want takers to have accumulated-value semantics.

val update_exn : ('a, Core_kernel.read_write) t -> f:('a -> 'a) -> unit: update_exn is like update, except it raises if is_empty t.

val read_only : ('a, [> Core_kernel.read ]) t -> ('a, Core_kernel.read) t

val write_only : ('a, [> Core_kernel.write ]) t -> ('a, Core_kernel.write) t

val value_available : (_, [> Core_kernel.read ]) t -> unit Async_kernel.Deferred.t

value_available t returns a deferred d that becomes determined when a value is in t. d does not include the value in t because that value may change after d becomes determined and before a deferred bind on d gets to run.

Repeated calls to value_available t will always return the same deferred until the t is filled.

val take : ('a, [> Core_kernel.read ]) t -> 'a Async_kernel.Deferred.t: take t returns a deferred that, when t is filled, becomes determined with the value of t and and clears t. If there are multiple concurrent calls to take then only one of them will be fulfilled and the others will continue waiting on future values. There is no ordering guarantee for which take call will be filled first.

val take_now : ('a, [> Core_kernel.read ]) t -> 'a option: take_now is an immediate form of take.

val take_now_exn : ('a, [> Core_kernel.read ]) t -> 'a
val taken : (_, [> Core_kernel.write ]) t -> unit Async_kernel.Deferred.t: taken t returns a deferred that is filled the next time take clears t.

val peek : ('a, [> Core_kernel.read ]) t -> 'a option: peek t returns the value in t without clearing t, or returns None is is_empty t.

val peek_exn : ('a, [> Core_kernel.read ]) t -> 'a: peek_exn t is like peek, except it raises if is_empty t.

val pipe_when_ready : ('a, [> Core_kernel.read ]) t -> 'a Async_kernel.Pipe.Reader.t

pipe_when_ready t returns a pipe, then repeatedly takes a value from t and writes it to the pipe. After each write, pipe_when_ready waits for the pipe to be ready to accept another value before taking the next value. Once the pipe is closed, pipe_when_ready will no longer take values from t.

Notice that this implementation effectively creates an extra buffer of size 1, so when you read from the pipe you can read a stale value (even though a fresh one should come immediately afterwards), and a value will be taken from the MVar even if it's never read from the pipe.

There is no protection against creating multiple pipes or otherwise multiple things trying to take concurrently. If that happens, it's not specified which of the pipes will get the value.