(define-module (8sync actors)
#:use-module (oop goops)
#:use-module (srfi srfi-9)
+ #:use-module (srfi srfi-9 gnu)
#:use-module (ice-9 control)
#:use-module (ice-9 format)
#:use-module (ice-9 match)
#:use-module (ice-9 atomic)
+ #:use-module ((ice-9 ports internal)
+ #:select (port-read-wait-fd port-write-wait-fd))
#:use-module (ice-9 pretty-print)
#:use-module (ice-9 receive)
+ #:use-module (ice-9 suspendable-ports)
#:use-module (fibers)
#:use-module (fibers channels)
#:use-module (fibers conditions)
#:use-module (fibers operations)
- #:use-module (8sync agenda)
#:use-module (8sync inbox)
#:use-module (8sync rmeta-slot)
+
#:export (;; utilities... ought to go in their own module
big-random-number
big-random-number-string
actor-id
actor-message-handler
+ *current-actor*
+
;;; Commenting out the <address> type for now;
;;; it may be back when we have better serializers
;; <address>
actor-init! actor-cleanup!
- actor-alive?
-
build-actions
define-actor
+ actor-spawn-fiber
+ with-actor-nonblocking-ports
+
;; <hive>
;; make-hive
;; ;; There are more methods for the hive, but there's
;; ;; no reason for the outside world to look at them maybe?
;; hive-id
- bootstrap-actor bootstrap-actor*
-
create-actor create-actor*
self-destruct
<- <-wait
- spawn-hive run-hive))
+ spawn-hive run-hive
+
+ ;; Maybe the wrong place for this, or for it to be exported.
+ ;; But it's used in websockets' server implementation at least...
+ live-wrap
+
+ *debug-actor-ids*))
;; For ids
(set! *random-state* (random-state-from-platform))
;; This is the internal, generalized message sending method.
;; Users shouldn't use it! Use the <-foo forms instead.
-(define-inlinable (%<- wants-reply from-actor to action args message-id in-reply-to)
+(define (%<- wants-reply from-actor to action args message-id in-reply-to)
;; Okay, we need to deal with message ids.
;; Could we get rid of them? :\
;; It seems if we can use eq? and have messages be immutable then
;; If we need to track replies across hive boundaries we could
;; register unique ids across the ambassador barrier.
(match to
- (#(_ _ (? channel? channel) dead?)
+ (($ <address> _ _ (? channel? channel) dead?)
(let ((message (make-message message-id to
(and from-actor (actor-id from-actor))
action args
(put-operation channel message)
(wait-operation dead?)))))
;; TODO: put remote addresses here.
- (#(actor-id hive-id #f #f)
+ (($ <address> actor-id hive-id #f #f)
;; Here we'd make a call to our hive...
'TODO)
;; A message sent to nobody goes nowhere.
;; TODO: Should we display a warning here, probably?
- (#f #f)))
+ (#f #f)
+ ;; We shouldn't technically be passing in actors but rather their
+ ;; addresses, but often actors want to message themselves and
+ ;; this makes that slightly easier.
+ ((? (lambda (x) (is-a? x <actor>)) actor)
+ (%<- wants-reply from-actor (actor-id actor) action
+ args message-id in-reply-to))))
(define (<- to action . args)
(define from-actor (*current-actor*))
(big-random-number-string))
#f))
-;; TODO: this should abort to the prompt, then check for errors
-;; when resuming.
-
(define (<-wait to action . args)
(define prompt (*actor-prompt*))
(when (not prompt)
(error "Tried to <-wait without being in an actor's context..."))
- (let ((reply (abort-to-prompt prompt to action args)))
- (cond ((eq? action '*error*)
+ (let ((reply (abort-to-prompt prompt '<-wait to action args)))
+ (cond ((eq? (message-action reply) '*error*)
(throw 'hive-unresumable-coroutine
"Won't resume coroutine; got an *error* as a reply"
#:message reply))
#:message message))
(apply method actor message (message-body message)))
+(define-syntax-rule (live-wrap body)
+ "Wrap possibly multi-value function in a procedure, applies all arguments"
+ (lambda args
+ (apply body args)))
+
(define-syntax-rule (build-actions (symbol method) ...)
"Construct an alist of (symbol . method), where the method is wrapped
-with wrap-apply to facilitate live hacking and allow the method definition
+with `live-wrap' to facilitate live hacking and allow the method definition
to come after class definition."
(build-rmeta-slot
(list (cons (quote symbol)
- (wrap-apply method)) ...)))
+ (live-wrap method)) ...)))
(define-class <actor> ()
- ;; An address object... a vector of #(actor-id hive-id inbox-channel dead?)
- ;; - inbox-channel is the receiving channel (as opposed to actor-inbox-deq)
- ;; - dead? is a fibers condition variable which is set once this actor
- ;; kicks the bucket
+ ;; An <address> object
(id #:init-keyword #:address
#:getter actor-id)
- ;; The connection to the hive we're connected to.
- (hive-channel #:init-keyword #:hive-channel
- #:accessor actor-hive-channel)
;; Our queue to send/receive messages on
(inbox-deq #:init-thunk make-channel
(define-method (actor-cleanup! (actor <actor>))
'no-op)
-;;; Addresses are vectors where the first part is the actor-id and
-;;; the second part is the hive-id. This works well enough... they
-;;; look decent being pretty-printed.
-
-(define (make-address actor-id hive-id channel dead?)
- (vector actor-id hive-id channel dead?))
-
-(define (address-actor-id address)
- (vector-ref address 0))
-
-(define (address-hive-id address)
- (vector-ref address 1))
-
-(define (address-channel address)
- (vector-ref address 2))
-
-(define (address-dead? address)
- (vector-ref address 3))
+;;; Every actor has an address, which is how its identified.
+;;; We also pack in some routing information.
+(define-record-type <address>
+ (make-address actor-id hive-id channel dead?)
+ address?
+ ;; Unique-to-this-actor-on-this-hive part
+ (actor-id address-actor-id)
+ ;; Unique identifier for the hive we're connected to.
+ ;; If we don't have a "direct" link to the other actor through
+ ;; a channel, we'll have to look up if our hive has a way to route
+ ;; to the other hive.
+ (hive-id address-hive-id)
+ ;; The receiving channel (as opposed to actor-inbox-deq)
+ (channel address-channel)
+ ;; A fibers condition variable which is set once this actor kicks
+ ;; the bucket
+ (dead? address-dead?))
+
+(set-record-type-printer!
+ <address>
+ (lambda (address port)
+ (format port "<address ~a ~a>"
+ (address->string address)
+ (if (address-channel address)
+ (string-append
+ ":local"
+ (if (atomic-box-ref
+ ((@@ (fibers conditions) condition-signalled?)
+ (address-dead? address)))
+ " :dead" ""))
+ ":remote"))))
(define (address->string address)
(string-append (address-actor-id address) "@"
This compares the actor-id and hive-id but ignores the channel and
dead? condition."
- (match address1
- (#(actor-id-1 hive-id-1 _ _)
- (match address2
- (#(actor-id-2 hive-id-2)
- (and (equal? actor-id-1 actor-id-2)
- (and (equal? hive-id-1 hive-id-2))))
- (_ #f)))
- (_ #f)))
+ (and (equal? (address-actor-id address1)
+ (address-actor-id address2))
+ (equal? (address-hive-id address1)
+ (address-hive-id address2))))
(define (actor-id-actor actor)
"Get the actor id component of the actor-id"
(define *actor-prompt*
(make-parameter #f))
+(define *resume-io-channel*
+ (make-parameter #f))
+
(define (actor-main-loop actor)
"Main loop of the actor. Loops around, pulling messages off its queue
and handling them."
(define dead?
(address-dead? (actor-id actor)))
(define prompt (make-prompt-tag (actor-id-actor actor)))
+ ;; Not always used, only if with-actor-nonblocking-ports is used
+ (define resume-io-channel
+ (make-channel))
(define (handle-message message)
(catch #t
(lambda vals
;; Return reply if necessary
(when (message-wants-reply message)
- (when (message-wants-reply message)
- (%<- #f actor (message-from message) '*reply*
- vals ((actor-msg-id-generator actor))
- (message-id message)))))))
+ (%<- #f actor (message-from message) '*reply*
+ vals ((actor-msg-id-generator actor))
+ (message-id message))))))
(const #t)
(let ((err (current-error-port)))
(lambda (key . args)
"Tried to resume nonexistant message: ~a\n"
(message-id message)))))
+ (define (call-with-actor-prompt thunk)
+ (call-with-prompt prompt
+ thunk
+ ;; Here's where we abort to if we're doing <-wait
+ ;; @@: maybe use match-lambda if we're going to end up
+ ;; handling multiple ~commands
+ (match-lambda*
+ ((kont '<-wait to action message-args)
+ (define message-id
+ ((actor-msg-id-generator actor)))
+ (hash-set! waiting message-id kont)
+ (%<- #t actor to action message-args message-id #f))
+ ((kont 'run-me proc)
+ (proc kont)))))
+
(define halt-or-handle-message
;; It would be nice if we could give priorities to certain operations.
;; halt should always win over getting a message...
(const #f)) ; halt and return
(wrap-operation (get-operation (actor-inbox-deq actor))
(lambda (message)
- (call-with-prompt prompt
- (lambda ()
- (if (message-in-reply-to message)
- ;; resume a continuation which was waiting on a reply
- (resume-handler message)
- ;; start handling a new message
- (handle-message message)))
- ;; Here's where we abort to if we're doing <-wait
- ;; @@: maybe use match-lambda if we're going to end up
- ;; handling multiple ~commands
- (lambda (kont to action message-args)
- (define message-id
- ((actor-msg-id-generator actor)))
- (hash-set! waiting message-id kont)
- (%<- #t actor to action message-args message-id #f)))
- #t)))) ; loop again
+ (call-with-actor-prompt
+ (lambda ()
+ (if (message-in-reply-to message)
+ ;; resume a continuation which was waiting on a reply
+ (resume-handler message)
+ ;; start handling a new message
+ (handle-message message))))
+ #t)) ; loop again
+ (wrap-operation (get-operation resume-io-channel)
+ (lambda (thunk)
+ (call-with-actor-prompt
+ (lambda ()
+ (thunk)))
+ #t))))
;; Mutate the parameter; this should be fine since each fiber
;; runs in its own dynamic state with with-dynamic-state.
;; See with-dynamic-state discussion in
;; https://wingolog.org/archives/2017/06/27/growing-fibers
(*current-actor* actor)
+ (*resume-io-channel* resume-io-channel)
+
;; We temporarily set the *actor-prompt* to #f to make sure that
;; actor-init! doesn't try to do a <-wait message (and not accidentally use
;; the parent fiber's *actor-prompt* either.)
(and (perform-operation halt-or-handle-message)
(loop))))
+
+;; @@: So in order for this to work, we're going to have to add
+;; another channel to actors, which is resumable i/o. We'll have to
+;; spawn a fiber that wakes up a thunk on the actor when its port is
+;; available. Funky...
+
+(define (%suspend-io-to-actor wait-for-read/write)
+ (lambda (port)
+ (define prompt (*actor-prompt*))
+ (define resume-channel (*resume-io-channel*))
+ (define (run-at-prompt k)
+ (spawn-fiber
+ (lambda ()
+ (wait-for-read/write port)
+ ;; okay, we're awake again, tell the actor to resume this
+ ;; continuation
+ (put-message resume-channel k))
+ #:parallel? #f))
+ (when (not prompt)
+ (error "Attempt to abort to actor prompt outside of actor"))
+ (abort-to-prompt (*actor-prompt*)
+ 'run-me run-at-prompt)))
+
+(define suspend-read-to-actor
+ (%suspend-io-to-actor (@@ (fibers) wait-for-readable)))
+
+(define suspend-write-to-actor
+ (%suspend-io-to-actor (@@ (fibers) wait-for-writable)))
+
+(define (with-actor-nonblocking-ports thunk)
+ "Runs THUNK in dynamic context in which attempting to read/write
+from a port that would otherwise block an actor's correspondence with
+other actors (note that reading from a nonblocking port should never
+block other fibers) will instead permit reading other messages while
+I/O is waiting to complete.
+
+Note that currently "
+ (parameterize ((current-read-waiter suspend-read-to-actor)
+ (current-write-waiter suspend-write-to-actor))
+ (thunk)))
+
+(define (actor-spawn-fiber thunk . args)
+ "Spawn a fiber from an actor but unset actor-machinery-specific
+dynamic context."
+ (apply spawn-fiber
+ (lambda ()
+ (*current-actor* #f)
+ (*resume-io-channel* #f)
+ (*actor-prompt* #f)
+ (thunk))
+ args))
+
+
\f
;;; Actor utilities
;;; ===============
;;; Every actor has a hive, which keeps track of other actors, manages
;;; cleanup, and performs inter-hive communication.
+;; TODO: Make this a srfi-9 record type
(define-class <hive> ()
(id #:init-keyword #:id
#:getter hive-id)
(and (perform-operation halt-or-handle)
(lp))))
-(define *current-hive* (make-parameter #f))
+(define *hive-id* (make-parameter #f))
+(define *hive-channel* (make-parameter #f))
+;; @@: Should we halt the hive either at the end of spawn-hive or run-hive?
(define* (spawn-hive proc #:key (hive (make-hive)))
- "Spawn a hive in a fiber running PROC, passing it the fresh hive"
+ "Spawn a hive and run PROC, passing it the fresh hive and establishing
+a dynamic context surrounding the hive."
(spawn-fiber (lambda () (hive-main-loop hive)))
- (proc hive))
+ (parameterize ((*hive-id* (hive-id hive))
+ (*hive-channel* (hive-channel hive)))
+ (proc hive)))
(define (run-hive proc . args)
"Spawn a hive and run it in run-fibers. Takes a PROC as would be passed
(spawn-hive proc))
args))
-(define (%create-actor hive-channel hive-id
- actor-class init-args id-cookie send-init?)
- (let* ((actor-id (gen-actor-id id-cookie))
+(define (%create-actor actor-class init-args id-cookie send-init?)
+ (let* ((hive-channel (*hive-channel*))
+ (hive-id (*hive-id*))
+ (actor-id (gen-actor-id id-cookie))
(dead? (make-condition))
(inbox-enq (make-channel))
(address (make-address actor-id hive-id
inbox-enq dead?))
(actor (apply make actor-class
- #:hive-channel hive-channel
#:address address
init-args))
(should-init (actor-should-init actor)))
;; return the address
address))
-(define* (bootstrap-actor hive actor-class #:rest init-args)
- "Create an actor on HIVE using ACTOR-CLASS passing in INIT-ARGS args"
- (%create-actor (hive-channel hive) (hive-id hive) actor-class
- init-args (symbol->string (class-name actor-class))
- #f))
-
-(define* (bootstrap-actor* hive actor-class id-cookie #:rest init-args)
- "Create an actor, but also allow customizing a 'cookie' added to the id
-for debugging"
- (%create-actor (hive-channel hive) (hive-id hive) actor-class
- init-args id-cookie
- #f))
+;;; Whether or not to attach the class' name as a cookie by default in
+;;; create-actor
+(define *debug-actor-ids*
+ (make-parameter #t))
-(define* (create-actor from-actor actor-class #:rest init-args)
+(define* (create-actor actor-class #:rest init-args)
"Create an instance of actor-class. Return the new actor's id.
This is the method actors should call directly (unless they want
to supply an id-cookie, in which case they should use
create-actor*)."
- (%create-actor (actor-hive-channel from-actor) (actor-id-hive from-actor)
- actor-class init-args #f #t))
+ (%create-actor actor-class init-args
+ (if (*debug-actor-ids*)
+ (symbol->string (class-name actor-class))
+ #f)
+ #t))
-
-(define* (create-actor* from-actor actor-class id-cookie #:rest init-args)
+(define* (create-actor* actor-class id-cookie #:rest init-args)
"Create an instance of actor-class. Return the new actor's id.
Like create-actor, but permits supplying an id-cookie."
- (%create-actor (actor-hive-channel from-actor) (actor-id-hive from-actor)
- actor-class init-args id-cookie #t))
+ (%create-actor actor-class init-args id-cookie #t))
-(define* (self-destruct actor #:key (cleanup #t))
+(define (self-destruct actor)
"Remove an actor from the hive.
-Unless #:cleanup is set to #f, this will first have the actor handle
-its '*cleanup* action handler."
+The actor will also call its `actor-cleanup!' method."
(signal-condition! (address-dead? (actor-id actor)))
- (put-message (actor-hive-channel actor) (list 'remove-actor (actor-id-actor actor)))
+ (put-message (*hive-channel*) (list 'remove-actor (actor-id-actor actor)))
;; Set *actor-prompt* to nothing to prevent actor-cleanup! from sending
;; a message with <-wait
(*actor-prompt* #f)
(actor-cleanup! actor))
-;; From a patch I sent to Fibers...
-(define (condition-signalled? cvar)
- "Return @code{#t} if @var{cvar} has already been signalled.
-
-In general you will want to use @code{wait} or @code{wait-operation} to
-wait on a condition. However, sometimes it is useful to see whether or
-not a condition has already been signalled without blocking."
- (atomic-box-ref ((@@ (fibers conditions) condition-signalled?) cvar)))
-
-(define (actor-alive? actor)
- (condition-signalled? (address-dead? (actor-id actor))))