Buffered output
Serialization primitives built for speed and memory-efficiency.
Buf_write is designed for writing fast and memory-efficient serializers. It is based on the Faraday library, but adapted for Eio. Its core type and related operation gives the user fine-grained control over copying and allocation behavior while serializing user-defined types, and presents the output in a form that makes it possible to use vectorized write operations, such as the writev system call, or any other platform or application-specific output APIs.
A Buf_write serializer manages an internal buffer and a queue of output buffers. The output bufferes may be a sub range of the serializer's internal buffer or one that is user-provided. Buffered writes such as string, char, cstruct, etc., copy the source bytes into the serializer's internal buffer. Unbuffered writes are done with schedule_cstruct, which performs no copying. Instead, it enqueues the source bytes into the serializer's write queue directly.
Example:
module Write = Eio.Buf_write
let () =
Eio_mock.Backend.run @@ fun () ->
let stdout = Eio_mock.Flow.make "stdout" in
Write.with_flow stdout (fun w ->
Write.string w "foo";
Write.string w "bar";
Eio.Fiber.yield ();
Write.string w "baz";
)This combines the first two writes, giving:
+stdout: wrote "foobar"
+stdout: wrote "baz"Raised when waiting for a flush to complete if the buffer is destroyed instead.
with_flow flow fn runs fn writer, where writer is a buffer that flushes to flow.
Concurrently with fn, it also runs a fiber that copies from writer to flow. If this fiber runs out of data to copy then it will suspend itself. Writing to writer will automatically schedule it to be resumed. This means that pending data is flushed automatically before the process sleeps.
When fn returns, writer is automatically closed and any remaining data is flushed before with_flow itself returns.
A serializer manages an internal buffer for coalescing small writes. The size of this buffer is determined when the serializer is created. If the buffer does not contain sufficient space to service a caller's buffered write, the serializer will allocate a new buffer of the sufficient size and use it for the current and subsequent writes. The old buffer will be garbage collected once all of its contents have been flushed.
val string : t -> ?off:int -> ?len:int -> string -> unitstring t ?off ?len str copies str into the serializer's internal buffer.
val bytes : t -> ?off:int -> ?len:int -> Stdlib.Bytes.t -> unitbytes t ?off ?len bytes copies bytes into the serializer's internal buffer. It is safe to modify bytes after this call returns.
val cstruct : t -> Cstruct.t -> unitcstruct t cs copies cs into the serializer's internal buffer. It is safe to modify cs after this call returns. For large cstructs, it may be more efficient to use schedule_cstruct.
val write_gen :
t ->
blit:('a -> src_off:int -> Cstruct.buffer -> dst_off:int -> len:int -> unit) ->
off:int ->
len:int ->
'a ->
unitwrite_gen t ~blit ~off ~len x copies x into the serializer's internal buffer using the provided blit operation. See Bigstring.blit for documentation of the arguments.
val char : t -> char -> unitchar t c copies c into the serializer's internal buffer.
val uint8 : t -> int -> unituint8 t n copies the lower 8 bits of n into the serializer's internal buffer.
module BE : sig ... endBig endian serializers
module LE : sig ... endLittle endian serializers
Unbuffered writes do not involve copying bytes to the serializer's internal buffer.
val schedule_cstruct : t -> Cstruct.t -> unitschedule_cstruct t cs schedules cs to be written. cs is not copied in this process, so cs should only be modified after t has been flushed.
val free_bytes_in_buffer : t -> intfree_bytes_in_buffer t returns the free space, in bytes, of the serializer's write buffer. If a write call has a length that exceeds this value, the serializer will allocate a new buffer that will replace the serializer's internal buffer for that and subsequent calls.
val has_pending_output : t -> boolhas_pending_output t is true if t's output queue is non-empty. It may be the case that t's queued output is being serviced by some other thread of control, but has not yet completed.
val pending_bytes : t -> intpending_bytes t is the size of the next write, in bytes, that t will surface to the caller via await_batch.
val pause : t -> unitpause t causes t to stop surfacing writes to the user. This gives the serializer an opportunity to collect additional writes before sending them to the underlying device, which will increase the write batch size.
As one example, code may want to call this function if it's about to release the OCaml lock and perform a blocking system call, but would like to batch output across that system call.
Call unpause to resume writing later. Note that calling flush or close will automatically call unpause too.
val flush : t -> unitflush t waits until all prior writes have been successfully completed. If t has no pending writes, flush returns immediately. If t is paused then it is unpaused first.
val close : t -> unitclose t closes t. All subsequent write calls will raise, and any subsequent pause calls will be ignored. If the serializer has any pending writes, user code will have an opportunity to service them before receiving End_of_file. Flush callbacks will continue to be invoked while output is shifted out of t as needed.
val is_closed : t -> boolis_closed t is true if close has been called on t and false otherwise. A closed t may still have pending output.
Low-level operations for running a serializer.
create ~sw len creates a serializer with a fixed-length internal buffer of length len. See the Buffered writes section for details about what happens when len is not large enough to support a write.
of_buffer ~sw buf creates a serializer, using buf as its internal buffer. The serializer takes ownership of buf until the serializer has been closed and flushed of all output.
val abort : t -> unitval await_batch : t -> Cstruct.t listawait_batch t returns a list of buffers that should be written. If no data is currently available, it waits until some is. After performing a write, call shift with the number of bytes written. You must accurately report the number of bytes written. Failure to do so will result in the same bytes being surfaced multiple times.
val shift : t -> int -> unitshift t n removes the first n bytes in t's write queue. Any flush operations called within this span of the write queue will be scheduled to resume.
These functions are included for testing, debugging, and general development. They are not the suggested way of driving a serializer in a production setting.
val serialize :
t ->
(Cstruct.t list -> (int, [ `Closed ]) Stdlib.result) ->
(unit, [> `Closed ]) Stdlib.resultserialize t writev calls writev bufs each time t is ready to write. In the event that writev indicates a partial write, serialize will call Fiber.yield before continuing.
val serialize_to_string : t -> stringserialize_to_string t runs t, collecting the output into a string and returning it. serializie_to_string t immediately closes t.
val serialize_to_cstruct : t -> Cstruct.tserialize_to_cstruct t runs t, collecting the output into a cstruct and returning it. serialize_to_cstruct t immediately closes t.
val drain : t -> intdrain t removes all pending writes from t, returning the number of bytes that were enqueued to be written and freeing any scheduled buffers in the process. Note that this does not close t itself, and does not return until t has been closed.