=encoding utf8 =head1 TITLE Synopsis 17: Concurrency [DRAFT] =head1 AUTHOR Elizabeth Mattijsen Audrey Tang =head1 VERSION Maintainer: Elizabeth Mattijsen Date: 13 Jun 2005 Last Modified: 13 Nov 2005 Number: 0 Version: 1 =head1 SKETCH This is a rough sketch of how concurrency works in Perl 6. (actually these are just random notes, put here under the release-early release-often principle, slowly being integrated in a more textual format. Patches welcome!) =head1 OVERVIEW Concurrency can take many forms in Perl 6. With varying degrees of explicitness and control capabilities. This document attempts to describe what these capabilities are and in which form they can be accessed in Perl 6. =head2 Processes, threads, fibers? Concurrency comes in many shapes and forms. Most Perl users are used to the concept of a "process" or a "thread" (usually depending on the OS they work on). Some systems even are familiar with very lightweight threads called "fibers". When discussing issues about concurrency with different people, it soon becomes apparent that everybody has his own set of "understandings" about what each word means, which doesn't make it any easier to describe Perl 6 concurrency. It seemed the most natural to use the word "thread" to describe a process which has its own context, but also shares context with 0 or more concurrently running processes. Depending on your OS, or even specific version of your OS, this could still be a single "process" from the OS's point of view. Or it could contain an OS process for each thread. Or any mixture of these two implementations. In this document we try to be agnostic about this: all we know in Perl 6 are "threads", which have their own context and share context with other concurrently running "threads". Whether they be process, threads or fibres at the OS level should not matter at the Perl 6 level. And for sake of consistency, an unthreaded "normal" program is considered to be also running in a single thread. =head2 Variables In the past, there have been two models for concurrent processes in Perl. In general, these are referred to as "5.005 threads" (C) and "ithreads" (C). The main difference between these two models from a programmer's point of view, is that variables in "5.005 threads" are shared by default. Whereas in the "ithreads" model, only variables that have been indicated to be "shared", are actually shared between threads. All other variable values are actually copies of the variable's value in the "parent" thread. With regards to variables, the concurrency model of Perl 6 is closer to the "5.005 threads" model than it is to the "ithreads" model. In fact, all variables "visible" to a particular scope in Perl 6 will be accessible and modifiable (if allowed to do so) from all of the concurrent processes that start from that scope. In that sense, one could consider the "ithreads" model as a historical diversion: the Perl 6 concurrency picks up where the "5.005 threads" path left off. (EM: maybe point out that the "ithreads" behaviour can be simulated with some kind of copy-on-write magic to be automagically added to all variable access inside a thread, except for those with an explicit "is shared" attribute?) =head2 No user accessible locks Differently from any current concurrent process implementation in Perl, there are no user accessible locks. Instead, the concept of Software Transactional Memory is used. This is in concept similar to the use of BEGIN TRANSACTION ... do your uninterruptible actions COMMIT in the database world. More interestingly, this also includes the concept of rollback: BEGIN TRANSACTION ... do your stuff, but impossible to complete: ROLLBACK This causes the state of the process to be reverted to the state at the moment the BEGIN TRANSACTION was executed. Perl 6 supports this concept through C blocks. These sections are guaranteed to either be completed totally (when the Code block is exited), or have their state reverted to the state at the start of the Code block (with the L statement). (EM: maybe point out if / how old style locks can be "simulated", for those needing a migration path?) =head2 Atomic Code blocks my ($x, $y); sub c { $x -= 3; $y += 3; $x < 10 or defer; } sub d { $x += 3; $y -= 3; $y < 10 or defer; } contend { # ... maybe { c() } maybe { d() }; # ... } A Code block can be prefixed with C. This means that code executed inside that scope is guaranteed not to be interrupted in any way. The start of a block marked C also becomes a I to which execution can return (in exactly the same state) if a problem occurs (a.k.a. a L is done) inside the scope of the Code block. =head3 defer The C function basically restores the state of the thread at the last checkpoint and will wait there until an external event allows it to potentially run that atomic C section of code again without having to defer again. If there are no external events possible that could restart execution, an exception will be raised. The last checkpoint is either the outermost C boundary, or the most immediate caller constructed with C. =head3 maybe The C statement causes a checkpoint to be made for C for each block in the C chain, creating an alternate execution path to be followed when a C is done. For example: maybe { ... some_condition() or defer; ... } maybe { ... some_other_condition() or defer; ... } maybe { ... } If placed outside a C block, the C statement creates its own C barrier. =head3 limitations Because Perl 6 must be able to revert its state to the state it had at the checkpoint, it is not allowed to perform any non-revertible actions. These would include reading / writing from file handles that do not support C (such as sockets). Attempting to do so will cause a fatal error to occur. This will probably need to be expanded to all objects: any object that has some interface with data "outside" of the knowledge of the language (e.g. an interface with an external XML library) would also need to provide some method for freezing a state, and restoring to a previously frozen state. If you're not interested in revertability, but are interested in uninterruptability, you could use the "is critical" trait. =head2 Critical Code blocks sub tricky is critical { # code accessing external info, not to be interrupted } if ($update) { is critical; # code accessing external info, not to be interrupted } A Code block marked "is critical" can not be interrupted in any way. But since it is able to access non-revertible data structures (such as non-seekable file handles), it cannot do a C as it would be impossible to restore the state to the beginning of the Code block. =head2 Mixing Atomic and Critical Both "atomic" as well as "critical" propagate down the call chain. This means that any subroutine that in itself is not "atomic" or "critical" becomes uninterruptible if called inside a code block that is marked as "atomic" or "critical". Atomic Code blocks called inside the call chain of a "critical" code block do not pose a problem, as they are more restrictive. Any code that attempts to perform any non-revertible action (e.g. reading from a socket) will cause a fatal error when called inside the call chain of an Atomic Code block. =head2 Co-Routines sub foo { my @list = ( 1, 11, 17 ); while (@list) { @list.shift.produce; } return; } say foo; # 1 say foo; # 11 say foo; # 17 say foo; # undef say foo; # 1 There is no real difference between a subroutine (or method) and a so-called 'co-routine'. The only difference is how control is relinquished by the subroutine. If this is done with a C statement, the subroutine is considered to be "normal" (i.e. the next time the subroutine is called, execution will start again at the top of the subroutine. A different situation occurs when the control is returned by a subroutine to its caller by means of a C statement. From the caller's point of view, there is no difference with the C statement. However, the next time the subroutine is called, execution will continue after the last C statement, B from the beginning of the subroutine. If there are no statements after the last C statement executed, then execution will start from the start of the called subroutine again. Please note that there is no concurrency involved with the C statement. Execution in the calling sub will halt until the called subroutine returns, regardless of whether this happens by a C or a C statement. Parameters passed to a subroutine that has previously returned with C will be handled as if if was an initial call to the subroutine. This means that the values of named parameters will be placed in their expected place inside the subroutine. Positional parameters will be available to any code looking at them, but will B be handled automatically. It would seem that the use of named parameters is therefore advisable. ####################################################################### Below here still the more or less unorganized stuff CORE::GLOBAL::exit; # kills all the threads # We intentionally do not list cross-machine parallelism Conc:: classes here. # Consult your local 6PAN mirror with a time machine. use Conc::Processes; # fork() or createProcess based implementation use Conc::Threads; # maybe it just exports &async to override the default one, yay use Conc::Multiplex; # this is default my $thr = async { ...do something... END { } }; Conc::Thread.this Conc::Proc.this Conc object # name is still up for grabs! - numify to TIDs (as in pugs) - stringify to something sensible (eg. ""); - enumerable with Conc.list - Conc.yield (if this is to live but deprecated, maybe call it sleep(0)?) - sleep() always respects other threads, thank you very much - standard methods: - .join # wait for invocant to finish (always item cxt) - .die # throw exception in the invocant thread - .alarm # set up alarms - .alarms # query existing alarms - .suspend # pause a thread; fail if already paused - .resume # revive a thread; fail if already running - .detach # survives parent thread demise (promoted to process) # process-local changes no longer affects parent # tentatively, the control methods still applies to it # including wait (which will always return undef) # also needs to discard any atomicity context - attributes: - .started # time - .finished # time - .waiting # suspended (not diff from block on wakeup signal) # waiting on a handle, a condition, a lock, et cetera # otherwise returns false for running threads # if it's finished then it's undef(?) - .current_continuation # the CC currently running in that thread - "is throttled" trait method throttled::trait_auxiliary: ($limit=1, :$key=gensym()) { # "is throttled" limits max connection to this Code object # the throttling is shared among closures with the same key # the limit may differ on closures with the same key. # if the counter with the "key" equals or exceeds a closure's limit, # the closure can't be entered until it's released # (this can be trivially implemented using contend+defer) } class Foo { method a is throttled(:limit(3) :key) { ... } method b is throttled(:limit(2) :key) { ... } } my Foo $f .= new; async { $f.a } async { $f.b } - Thread::Status - IO objects and containers gets concurrency love! - $obj.wake_on_readable - $obj.wake_on_writable - $obj.wake_on_either_readable_or_writable_or_passed_time(3); # fixme fixme - $obj.wake_on:{.readable} # busy wait, probably my @a is Array::Chan = 1..Inf; async { @a.push(1) }; async { @a.blocking_shift({ ... }) }; async { @a.unshift({ ... }) }; Communication abstractions - shared, transactional variables by default # program will wait for _all_ threads # unjoined threads will be joined at the beginning of the END block batch # of the parent thread that spawned them ### INTERFACE BARRIER ### module Blah; { is atomic; # contend/maybe/whatever other rollback stuff # limitation: no external IO (without lethal warnings anyway) # can't do anything irreversible is critical; # free to do anything irreversible # means "don't interrupt me" # in system with critical section, no interrupts from # other threads will happen during execution # you can't suspend me my $boo is export; $boo = 1; # We decree that this part forms the static interface # it's run once during initial compilation under the # Separate Compilation doctrine and the syms sealed off # to form part of bytecode syms headers %CALLER::<&blah> = { 1 }; # work - adds to export set die "Eureka!" if %CALLER::<$sym>; # never dies # BEGIN { $boo = time }; sub IMPORT { # VERY DYNAMIC! our $i = time; %CALLER::<&blah> = { 1 }; # work - adds to export set die "Eureka!" if %CALLER::<$sym>; # probes interactively } } ### INTERFACE BARRIER ### my $sym; threads.new({ use Blah; BEGIN { require(Blah).import } my $boo; BEGIN { eval slurp; $boo := $Blah::boo }; ... }); =head2 Signals Asynchronous exceptions are just like user-initiated exceptions with C, so you can also catch it with regular C blocks as specified in S04. To declare your main program catches INT signals, put a CATCH block anywhere in the toplevel to handle exceptions like this: CATCH { when Error::Signal::INT { ... } } =head2 Alarm An alarm is just a pre-arranged exception to be delivered to your program. By the time alarm has arrived, the current block may have already finished executing, so you would need to set up CATCH blocks in places where an alarm can rise to handle it properly. You can request an alarm using the number of seconds, or with a target date. It returns a proxy alarm object that you can do interesting things with. multi Alarm *alarm (Num $seconds = $CALLER::_, &do = {die Sig::ALARM}, :$repeat = 1) multi Alarm *alarm (Date $date, &do = {die Sig::ALARM}, :$repeat = 1) Perl 6's C has three additional features over traditional alarms: =head3 Multiple and Lexical Alarms One can set up multiple alarms using repeated alarm calls: { my $a1 = alarm(2); my $a2 = alarm(2); sleep 10; CATCH { is critical; # if you don't want $a2 to be raised inside this when Sig::ALARM { ... } } } To stop an alarm, call C<$alarm.stop>. The C method for Conc objects (including process and threads) returns a list of alarms currently scheduled for that concurrent context. When an alarm object is garbage collected, the alarm is stopped automatically. Under void context, the implicit alarm object can only be stopped by querying C<.alarms> on the current process. We are not sure what C would mean. Probably a deprecation warning? =head3 Repeated Alarms If you request a repeated alarm using the C named argument, it will attempt to fire off the alarm that many times. However, the alarm will be suppressed when inside a C block that's already handling the exception raised by I alarm. To repeat 0 times is to not fire off any alarms at all. To repeat +Inf times is to repeat over and over again. =head3 Callbacks in Alarms You can arrange a callback (like JavaScript's setTimeOut) in C, which will then be invoked with the then-current code as caller. If you set up such a callback to another Conc object, what happens is just like when you called C<.die> on behalf of that object -- namely, the callback closure, along with anything it referenced, is shared to the target Conc context. Unlike in Perl 5's ithreads where you cannot share anything after the fact, this allows passing shared objects in an C fashion across concurrent parts of the program. Under the default (multiplexing) concurrency model, this is basically a no-op. =head2 Continuations =head3 Coroutines ## braindump of coro meeting by Liz and Autri, more to follow - Coros are _like_ processes coro dbl { yield $_ * 2; yield $_; return }; my @x = 1..10; my %y = map &dbl, @x; # 2 => 2, 6 => 4, 10 => 6, ... coro perm (@x) { @x.splice(rand(@x),1).yield while @x; } my &p1 := &perm.start(1..10); my &p2 := &perm.start(1..20); p1(); p1(); p2(); p2(); coro foo { yield 42 }; (1..10).pick; coro foo ($x) { yield $x; yield $x+2; cleanup(); while (2) { while (1) { &?SUB.kill; # seppuku } } } # implicit falloff return + return() means start over without yielding # return() means yielding and restart + no implicit falloff (I LIKE THIS) &foo.finished; # true on return() and false on midway yield() foo(4); # and that's all she wrote coro foo ($x) { yield $x; # this point with $x bound to 10 yield $x+1; return 5; ... # this is never reached, I think we all agree } # If you don't want your variables to get rebound, use "is copy": coro foo ($x is copy) {...} # which is sugar for coro foo ($x) { { my $x := $OUTER::x; ...; # Further calls of &foo rebound $OUTER::x, not $x. } } sub foo { return undef if rand; ... } use overload { '&{}' => sub { ... } } class Coro is Conc::Multiplex does Code { method postcircumfix:<( )> { # start the thread, block stuff (we are in the caller's context) } } class Hash is extended { method postcircumfix:<( )> (&self: *@_) { &self = self.start(@_); } method start { # remember self # upon return() or normal falloff, restore self } } %*ENV(123); &foo_continued := &foo.start(10); &foo.start(20); foo(10); # returns 10 foo(); # be "insufficient param" error or just return 11? foo(20); # returns 21 # continuation coros multi foo () { ...no rebinding... } multi foo ($x) { ...rebinding... } &foo.kill; my $first_ret = zoro( type => ); &zoro.variant(:type).kill; &zoro.variant(type => 'even').kill; zoro( type => ); zoro( even => 1 ); zoro( odd => 1 ); multi coro zoro ($type where 'even') {} multi coro zoro ($type where 'odd') {} multi coro zoro ($even is named) {} multi coro zoro ($odd is named) {} # iblech's thoughts: # Coroutine parameters should never be rebound. Instead, yield(...)s return # value is an Arglist object containing the new arguments: coro bar ($a, $b) { ...; my $new_set_of_args = yield(...); my $sum_of_old_a_and_new_a = $a + $new_set_of_args<$a>; ...; } bar(42, 23); # $a is 42, $b is 23 bar(17, 19); # $a still 42, $b still 19, # $new_set_of_args is \(a => 17, b => 19) =head2 Junctive Autothreading and Hyper Operations Live in userland for the time being. =head2 Interprocess Communication =head2 I/O Considerations =head3 File Descriptors =head3 Sockets =cut