Coro::State - first class continuations


 use Coro::State;

 $new = new Coro::State sub {
    print "in coro (called with @_), switching back\n";
    $new->transfer ($main);
    print "in coro again, switching back\n";
    $new->transfer ($main);
 }, 5;

 $main = new Coro::State;

 print "in main, switching to coro\n";
 $main->transfer ($new);
 print "back in main, switch to coro again\n";
 $main->transfer ($new);
 print "back in main\n";


This module implements coro. Coros, similar to threads and continuations, allow you to run more than one "thread of execution" in parallel. Unlike so-called "kernel" threads, there is no parallelism and only voluntary switching is used so locking problems are greatly reduced. The latter is called "cooperative" threading as opposed to "preemptive" threading.

This can be used to implement non-local jumps, exception handling, continuation objects and more.

This module provides only low-level functionality. See Coro and related modules for a higher level threads abstraction including a scheduler.


Coro::State implements two different thread models: Perl and C. The C threads (called cctx's) are basically simplified perl interpreters running/interpreting the Perl threads. A single interpreter can run any number of Perl threads, so usually there are very few C threads.

When Perl code calls a C function (e.g. in an extension module) and that C function then calls back into Perl or transfers control to another thread, the C thread can no longer execute other Perl threads, so it stays tied to the specific thread until it returns to the original Perl caller, after which it is again available to run other Perl threads.

The main program always has its own "C thread" (which really is *the* Perl interpreter running the whole program), so there will always be at least one additional C thread. You can use the debugger (see Coro::Debug) to find out which threads are tied to their cctx and which aren't.


A newly created Coro::State that has not been used only allocates a relatively small (a hundred bytes) structure. Only on the first transfer will perl allocate stacks (a few kb, 64 bit architetcures use twice as much, i.e. a few kb :) and optionally a C stack/thread (cctx) for threads that recurse through C functions. All this is very system-dependent. On my x86-pc-linux-gnu system this amounts to about 2k per (non-trivial but simple) Coro::State.

You can view the actual memory consumption using Coro::Debug. Keep in mind that a for loop or other block constructs can easily consume 100-200 bytes per nesting level.



This works similarly to $SIG{__DIE__} and is used as the default die hook for newly created Coro::States. This is useful if you want some generic logging function that works for all threads that don't set their own hook.

When Coro::State is first loaded it will install these handlers for the main program, too, unless they have been overwritten already.

The default handlers provided will behave like the built-in ones (as if they weren't there).

If you don't want to exit your program on uncaught exceptions, you must not return from your die hook - call Coro::terminate instead.

Note 1: You must store a valid code reference in these variables, undef will not do.

Note 2: The value of this variable will be shared among all threads, so changing its value will change it in all threads that don't have their own die handler.


Similar to above die hook, but augments $SIG{__WARN__}.


$coro = new Coro::State [$coderef[, @args...]]

Create a new Coro::State thread object and return it. The first transfer call to this thread will start execution at the given coderef, with the given arguments.

Note that the arguments will not be copied. Instead, as with normal function calls, the thread receives passed arguments by reference, so make sure you don't change them in unexpected ways.

Returning from such a thread is NOT supported. Neither is calling exit or throwing an uncaught exception. The following paragraphs describe what happens in current versions of Coro.

If the subroutine returns the program will be terminated as if execution of the main program ended.

If it throws an exception the program will terminate unless the exception is caught, exactly like in the main program.

Calling exit in a thread does the same as calling it in the main program, but due to libc bugs on many BSDs, this doesn't work reliable everywhere.

If the coderef is omitted this function will create a new "empty" thread, i.e. a thread that cannot be transfered to but can be used to save the current thread state in (note that this is dangerous, as no reference is taken to ensure that the "current thread state" survives, the caller is responsible to ensure that the cloned state does not go away).

The returned object is an empty hash which can be used for any purpose whatsoever, for example when subclassing Coro::State.

Certain variables are "localised" to each thread, that is, certain "global" variables are actually per thread. Not everything that would sensibly be localised currently is, and not everything that is localised makes sense for every application, and the future might bring changes.

The following global variables can have different values per thread, and have the stated initial values:

   Variable       Initial Value
   @_             whatever arguments were passed to the Coro
   $_             undef
   $@             undef
   $/             "\n"
   $SIG{__DIE__}  aliased to $Coro::State::DIEHOOK(*)
   $SIG{__WARN__} aliased to $Coro::State::WARNHOOK(*)
   (default fh)   *STDOUT
   $^H, %^H       zero/empty.
   $1, $2...      all regex results are initially undefined

   (*) reading the value from %SIG is not supported, but local'ising is.

If you feel that something important is missing then tell me. Also remember that every function call that might call transfer (such as Coro::Channel::put) might clobber any global and/or special variables. Yes, this is by design ;) You can always create your own process abstraction model that saves these variables.

The easiest way to do this is to create your own scheduling primitive like in the code below, and use it in your threads:

  sub my_cede {
     local ($;, ...);

Another way is to use dynamic winders, see Coro::on_enter and Coro::on_leave for this.

$prev->transfer ($next)

Save the state of the current subroutine in $prev and switch to the thread saved in $next.

The "state" of a subroutine includes the scope, i.e. lexical variables and the current execution state (subroutine, stack).


Returns true iff this Coro::State object is "new", i.e. has never been run yet. Those states basically consist of only the code reference to call and the arguments, but consumes very little other resources. New states will automatically get assigned a perl interpreter when they are transfered to.


Returns true iff the Coro::State object has been destroyed (by cancel), i.e. it's resources freed because they were cancel'd (or terminate'd).


Forcefully destructs the given Coro::State. While you can keep the reference, and some memory is still allocated, the Coro::State object is effecticely dead, destructors have been freed, it cannot be transfered to anymore.

$state->throw ([$scalar])

See Coro->throw.

$state->call ($coderef)

Try to call the given $coderef in the context of the given state. This works even when the state is currently within an XS function, and can be very dangerous. You can use it to acquire stack traces etc. (see the Coro::Debug module for more details). The coderef MUST NOT EVER transfer to another state.

$state->eval ($string)

Like call, but eval's the string. Dangerous.


Swap the current $_ (swap_defsv) or @_ (swap_defav) with the equivalent in the saved state of $state. This can be used to give the coro a defined content for @_ and $_ before transfer'ing to it.

$state->swap_sv (\$sv, \$swap_sv)

This (very advanced) function can be used to make any variable local to a thread.

It works by swapping the contents of $sv and $swap_sv each time the thread is entered and left again, i.e. it is similarly to:

   $tmp = $sv; $sv = $swap_sv; $swap_sv = $tmp;

Except that it doesn't make an copies and works on hashes and even more exotic values (code references!).

Needless to say, this function can be very very dangerous: you can easily swap a hash with a reference (i.e. %hash becomes a reference), and perl will not like this at all.

It will also swap "magicalness" - so when swapping a builtin perl variable (such as $.), it will lose it's magicalness, which, again, perl will not like, so don't do it.

Lastly, the $swap_sv itself will be used, not a copy, so make sure you give each thread it's own $swap_sv instance.

It is, however, quite safe to swap some normal variable with another. For example, PApp::SQL stores the default database handle in $PApp::SQL::DBH. To make this a per-thread variable, use this:

   my $private_dbh = ...;
   $coro->swap_sv (\$PApp::SQL::DBH, \$private_dbh);

This results in $PApp::SQL::DBH having the value of $private_dbh while it executes, and whatever other value it had when it doesn't execute.

You can also swap hashes and other values:

   my %private_hash;
   $coro->swap_sv (\%some_hash, \%private_hash);
$state->trace ($flags)

Internal function to control tracing. I just mention this so you can stay away from abusing it.

$bytes = $state->rss

Returns the memory allocated by the coro (which includes static structures, various perl stacks but NOT local variables, arguments or any C context data). This is a rough indication of how much memory it might use.


Returns whether the state currently uses a cctx/C context. An active state always has a cctx, as well as the main program. Other states only use a cctxts when needed.


Forces the allocation of a C context for the currently running coro (if not already done). Apart from benchmarking there is little point in doing so, however.

$ncctx = Coro::State::cctx_count

Returns the number of C-level coro allocated. If this number is very high (more than a dozen) it might help to identify points of C-level recursion in your code and moving this into a separate coro.

$nidle = Coro::State::cctx_idle

Returns the number of allocated but idle (free for reuse) C level coro. Currently, Coro will limit the number of idle/unused cctxs to 8.

$old = Coro::State::cctx_stacksize [$new_stacksize]

Returns the current C stack size and optionally sets the new minimum stack size to $new_stacksize longs. Existing stacks will not be changed, but Coro will try to replace smaller stacks as soon as possible. Any Coro::State that starts to use a stack after this call is guaranteed this minimum stack size.

Please note that coros will only need to use a C-level stack if the interpreter recurses or calls a function in a module that calls back into the interpreter, so use of this feature is usually never needed.

$old = Coro::State::cctx_max_idle [$new_count]

Coro caches C contexts that are not in use currently, as creating them from scratch has some overhead.

This function returns the current maximum number of idle C contexts and optionally sets the new amount. The count must be at least 1, with the default being 4.

@states = Coro::State::list

Returns a list of all states currently allocated.

$was_enabled = Coro::State::enable_times [$enable]

Enables/disables/queries the current state of per-thread real and cpu-time gathering.

When enabled, the real time and the cpu time (user + system time) spent in each thread is accumulated. If disabled, then the accumulated times will stay as they are (they start at 0).

Currently, cpu time is only measured on GNU/Linux systems, all other systems only gather real time.

Enabling time profiling slows down thread switching by a factor of 2 to 10, depending on platform on hardware.

The times will be displayed when running Coro::Debug::command "ps", and cna be queried by calling $state->times.

($real, $cpu) = $state->times

Returns the real time and cpu times spent in the given $state. See Coro::State::enable_times for more info.

$clone = $state->clone

This exciting method takes a Coro::State object and clones it, i.e., it creates a copy. This makes it possible to restore a state more than once, and even return to states that have returned or have been terminated.

Since its only known purpose is for intellectual self-gratification, and because it is a difficult piece of code, it is not enabled by default, and not supported.

Here are a few little-known facts: First, coros *are* full/true/real continuations. Secondly Coro::State objects (without clone) *are* first class continuations. Thirdly, nobody has ever found a use for the full power of call/cc that isn't better (faster, easier, more efficiently) implemented differently, and nobody has yet found a useful control construct that can't be implemented without it already, just much faster and with fewer resources. And lastly, Scheme's call/cc doesn't support using call/cc to implement threads.

Among the games you can play with this is implementing a scheme-like call-with-current-continuation, as the following code does (well, with small differences).

   # perl disassociates from local lexicals on frame exit,
   # so use a global variable for return values.
   my @ret;

   sub callcc($@) {
      my ($func, @arg) = @_;

      my $continuation = new Coro::State;
      $continuation->transfer (new Coro::State sub {
         my $escape = sub {
            @ret = @_;
            Coro::State->new->transfer ($continuation->clone);
         $escape->($func->($escape, @arg));

      my @ret_ = @ret; @ret = ();
      wantarray ? @ret_ : pop @ret_

Which could be used to implement a loop like this:

   async {
      my $n; 
      my $l = callcc sub { $_[0] };
      print "iteration $n\n";

      $l->($l) unless $n == 10;

If you find this confusing, then you already understand the coolness of call/cc: It can turn anything into spaghetti code real fast.

Besides, call/cc is much less useful in a Perl-like dynamic language (with references, and its scoping rules) then in, say, scheme.

Now, the known limitations of clone:

It probably only works on perl 5.10; it cannot clone a coro inside the substition operator (but windows perl can't fork from there either) and some other contexts, and abort () is the preferred mechanism to signal errors. It cannot clone a state that has a c context attached (implementing clone on the C level is too hard for me to even try), which rules out calling call/cc from the main coro. It cannot clone a context that hasn't even been started yet. It doesn't work with -DDEBUGGING (but what does). It probably also leaks, and sometimes triggers a few assertions inside Coro. Most of these limitations *are* fixable with some effort, but that's pointless just to make a point that it could be done.

The current implementation could without doubt be optimised to be a constant-time operation by doing lazy stack copying, if somebody were insane enough to invest the time.


This module is not thread-safe. You must only ever use this module from the same thread (this requirement might be removed in the future).




 Marc Lehmann <>