--- Coro/Coro.pm 2009/06/17 21:38:23 1.256 +++ Coro/Coro.pm 2011/05/12 23:24:28 1.296 @@ -42,14 +42,15 @@ thread models. Unlike the so-called "Perl threads" (which are not actually real threads -but only the windows process emulation ported to unix, and as such act -as processes), Coro provides a full shared address space, which makes -communication between threads very easy. And Coro's threads are fast, -too: disabling the Windows process emulation code in your perl and using -Coro can easily result in a two to four times speed increase for your -programs. A parallel matrix multiplication benchmark runs over 300 times -faster on a single core than perl's pseudo-threads on a quad core using -all four cores. +but only the windows process emulation (see section of same name for +more details) ported to UNIX, and as such act as processes), Coro +provides a full shared address space, which makes communication between +threads very easy. And coro threads are fast, too: disabling the Windows +process emulation code in your perl and using Coro can easily result in +a two to four times speed increase for your programs. A parallel matrix +multiplication benchmark (very communication-intensive) runs over 300 +times faster on a single core than perls pseudo-threads on a quad core +using all four cores. Coro achieves that by supporting multiple running interpreters that share data, which is especially useful to code pseudo-parallel processes and @@ -65,12 +66,262 @@ See also the C section at the end of this document - the Coro module family is quite large. +=head1 CORO THREAD LIFE CYCLE + +During the long and exciting (or not) life of a coro thread, it goes +through a number of states: + +=over 4 + +=item 1. Creation + +The first thing in the life of a coro thread is it's creation - +obviously. The typical way to create a thread is to call the C function: + + async { + # thread code goes here + }; + +You can also pass arguments, which are put in C<@_>: + + async { + print $_[1]; # prints 2 + } 1, 2, 3; + +This creates a new coro thread and puts it into the ready queue, meaning +it will run as soon as the CPU is free for it. + +C will return a coro object - you can store this for future +reference or ignore it, the thread itself will keep a reference to it's +thread object - threads are alive on their own. + +Another way to create a thread is to call the C constructor with a +code-reference: + + new Coro sub { + # thread code goes here + }, @optional_arguments; + +This is quite similar to calling C, but the important difference is +that the new thread is not put into the ready queue, so the thread will +not run until somebody puts it there. C is, therefore, identical to +this sequence: + + my $coro = new Coro sub { + # thread code goes here + }; + $coro->ready; + return $coro; + +=item 2. Startup + +When a new coro thread is created, only a copy of the code reference +and the arguments are stored, no extra memory for stacks and so on is +allocated, keeping the coro thread in a low-memory state. + +Only when it actually starts executing will all the resources be finally +allocated. + +The optional arguments specified at coro creation are available in C<@_>, +similar to function calls. + +=item 3. Running / Blocking + +A lot can happen after the coro thread has started running. Quite usually, +it will not run to the end in one go (because you could use a function +instead), but it will give up the CPU regularly because it waits for +external events. + +As long as a coro thread runs, it's coro object is available in the global +variable C<$Coro::current>. + +The low-level way to give up the CPU is to call the scheduler, which +selects a new coro thread to run: + + Coro::schedule; + +Since running threads are not in the ready queue, calling the scheduler +without doing anything else will block the coro thread forever - you need +to arrange either for the coro to put woken up (readied) by some other +event or some other thread, or you can put it into the ready queue before +scheduling: + + # this is exactly what Coro::cede does + $Coro::current->ready; + Coro::schedule; + +All the higher-level synchronisation methods (Coro::Semaphore, +Coro::rouse_*...) are actually implemented via C<< ->ready >> and C<< +Coro::schedule >>. + +While the coro thread is running it also might get assigned a C-level +thread, or the C-level thread might be unassigned from it, as the Coro +runtime wishes. A C-level thread needs to be assigned when your perl +thread calls into some C-level function and that function in turn calls +perl and perl then wants to switch coroutines. This happens most often +when you run an event loop and block in the callback, or when perl +itself calls some function such as C or methods via the C +mechanism. + +=item 4. Termination + +Many threads actually terminate after some time. There are a number of +ways to terminate a coro thread, the simplest is returning from the +top-level code reference: + + async { + # after returning from here, the coro thread is terminated + }; + + async { + return if 0.5 < rand; # terminate a little earlier, maybe + print "got a chance to print this\n"; + # or here + }; + +Any values returned from the coroutine can be recovered using C<< ->join +>>: + + my $coro = async { + "hello, world\n" # return a string + }; + + my $hello_world = $coro->join; + + print $hello_world; + +Another way to terminate is to call C<< Coro::terminate >>, which at any +subroutine call nesting level: + + async { + Coro::terminate "return value 1", "return value 2"; + }; + +And yet another way is to C<< ->cancel >> (or C<< ->safe_cancel >>) the +coro thread from another thread: + + my $coro = async { + exit 1; + }; + + $coro->cancel; # also accepts values for ->join to retrieve + +Cancellation I be dangerous - it's a bit like calling C without +actually exiting, and might leave C libraries and XS modules in a weird +state. Unlike other thread implementations, however, Coro is exceptionally +safe with regards to cancellation, as perl will always be in a consistent +state, and for those cases where you want to do truly marvellous things +with your coro while it is being cancelled - that is, make sure all +cleanup code is executed from the thread being cancelled - there is even a +C<< ->safe_cancel >> method. + +So, cancelling a thread that runs in an XS event loop might not be the +best idea, but any other combination that deals with perl only (cancelling +when a thread is in a C method or an C for example) is +safe. + +=item 5. Cleanup + +Threads will allocate various resources. Most but not all will be returned +when a thread terminates, during clean-up. + +Cleanup is quite similar to throwing an uncaught exception: perl will +work it's way up through all subroutine calls and blocks. On it's way, it +will release all C variables, undo all C's and free any other +resources truly local to the thread. + +So, a common way to free resources is to keep them referenced only by my +variables: + + async { + my $big_cache = new Cache ...; + }; + +If there are no other references, then the C<$big_cache> object will be +freed when the thread terminates, regardless of how it does so. + +What it does C do is unlock any Coro::Semaphores or similar +resources, but that's where the C methods come in handy: + + my $sem = new Coro::Semaphore; + + async { + my $lock_guard = $sem->guard; + # if we reutrn, or die or get cancelled, here, + # then the semaphore will be "up"ed. + }; + +The C function comes in handy for any custom cleanup you +might want to do: + + async { + my $window = new Gtk2::Window "toplevel"; + # The window will not be cleaned up automatically, even when $window + # gets freed, so use a guard to ensure it's destruction + # in case of an error: + my $window_guard = Guard::guard { $window->destroy }; + + # we are safe here + }; + +Last not least, C can often be handy, too, e.g. when temporarily +replacing the coro thread description: + + sub myfunction { + local $Coro::current->{desc} = "inside myfunction(@_)"; + + # if we return or die here, the description will be restored + } + +=item 6. Viva La Zombie Muerte + +Even after a thread has terminated and cleaned up it's resources, the coro +object still is there and stores the return values of the thread. Only in +this state will the coro object be "reference counted" in the normal perl +sense: the thread code keeps a reference to it when it is active, but not +after it has terminated. + +The means the coro object gets freed automatically when the thread has +terminated and cleaned up and there arenot other references. + +If there are, the coro object will stay around, and you can call C<< +->join >> as many times as you wish to retrieve the result values: + + async { + print "hi\n"; + 1 + }; + + # run the async above, and free everything before returning + # from Coro::cede: + Coro::cede; + + { + my $coro = async { + print "hi\n"; + 1 + }; + + # run the async above, and clean up, but do not free the coro + # object: + Coro::cede; + + # optionally retrieve the result values + my @results = $coro->join; + + # now $coro goes out of scope, and presumably gets freed + }; + +=back + =cut package Coro; -use strict qw(vars subs); -no warnings "uninitialized"; +use common::sense; + +use Carp (); use Guard (); @@ -82,9 +333,9 @@ our $main; # main coro our $current; # current coro -our $VERSION = 5.132; +our $VERSION = 5.372; -our @EXPORT = qw(async async_pool cede schedule terminate current unblock_sub); +our @EXPORT = qw(async async_pool cede schedule terminate current unblock_sub rouse_cb rouse_wait); our %EXPORT_TAGS = ( prio => [qw(PRIO_MAX PRIO_HIGH PRIO_NORMAL PRIO_LOW PRIO_IDLE PRIO_MIN)], ); @@ -125,38 +376,25 @@ usually better to rely on L or L, as this is pretty low-level functionality. -This variable stores either a Coro object or a callback. +This variable stores a Coro object that is put into the ready queue when +there are no other ready threads (without invoking any ready hooks). -If it is a callback, the it is called whenever the scheduler finds no -ready coros to run. The default implementation prints "FATAL: -deadlock detected" and exits, because the program has no other way to -continue. - -If it is a coro object, then this object will be readied (without -invoking any ready hooks, however) when the scheduler finds no other ready -coros to run. +The default implementation dies with "FATAL: deadlock detected.", followed +by a thread listing, because the program has no other way to continue. This hook is overwritten by modules such as C and -C to wait on an external event that hopefully wake up a +C to wait on an external event that hopefully wakes up a coro so the scheduler can run it. -Note that the callback I, under any circumstances, block -the current coro. Normally, this is achieved by having an "idle -coro" that calls the event loop and then blocks again, and then -readying that coro in the idle handler, or by simply placing the idle -coro in this variable. - -See L or L for examples of using this -technique. - -Please note that if your callback recursively invokes perl (e.g. for event -handlers), then it must be prepared to be called recursively itself. +See L or L for examples of using this technique. =cut -$idle = sub { - require Carp; - Carp::croak ("FATAL: deadlock detected"); +# ||= because other modules could have provided their own by now +$idle ||= new Coro sub { + require Coro::Debug; + die "FATAL: deadlock detected.\n" + . Coro::Debug::ps_listing (); }; # this coro is necessary because a coro @@ -166,7 +404,7 @@ $manager = new Coro sub { while () { - Coro::State::cancel shift @destroy + _destroy shift @destroy while @destroy; &schedule; @@ -209,14 +447,6 @@ print "@_\n"; } 1,2,3,4; -=cut - -sub async(&@) { - my $coro = new Coro @_; - $coro->ready; - $coro -} - =item async_pool { ... } [@args...] Similar to C, but uses a coro pool, so you should not call @@ -282,7 +512,7 @@ Calls the scheduler. The scheduler will find the next coro that is to be run from the ready queue and switches to it. The next coro to be run is simply the one with the highest priority that is longest -in its ready queue. If there is no coro ready, it will clal the +in its ready queue. If there is no coro ready, it will call the C<$Coro::idle> hook. Please note that the current coro will I be put into the ready @@ -318,7 +548,8 @@ =item terminate [arg...] -Terminates the current coro with the given status values (see L). +Terminates the current coro with the given status values (see +L). The values will not be copied, but referenced directly. =item Coro::on_enter BLOCK, Coro::on_leave BLOCK @@ -502,22 +733,78 @@ =item $coro->cancel (arg...) -Terminates the given Coro and makes it return the given arguments as -status (default: the empty list). Never returns if the Coro is the +Terminates the given Coro thread and makes it return the given arguments as +status (default: an empty list). Never returns if the Coro is the current Coro. -=cut - -sub cancel { - my $self = shift; +This is a rather brutal way to free a coro, with some limitations - if +the thread is inside a C callback that doesn't expect to be canceled, +bad things can happen, or if the cancelled thread insists on running +complicated cleanup handlers that rely on it'S thread context, things will +not work. + +Any cleanup code being run (e.g. from C blocks) will be run without +a thread context, and is not allowed to switch to other threads. On the +plus side, C<< ->cancel >> will always clean up the thread, no matter +what. If your cleanup code is complex or you want to avoid cancelling a +C-thread that doesn't know how to clean up itself, it can be better to C<< +->throw >> an exception, or use C<< ->safe_cancel >>. + +The arguments to C<< ->cancel >> are not copied, but instead will +be referenced directly (e.g. if you pass C<$var> and after the call +change that variable, then you might change the return values passed to +e.g. C, so don't do that). + +The resources of the Coro are usually freed (or destructed) before this +call returns, but this can be delayed for an indefinite amount of time, as +in some cases the manager thread has to run first to actually destruct the +Coro object. + +=item $coro->safe_cancel ($arg...) + +Works mostly like C<< ->cancel >>, but is inherently "safer", and +consequently, can fail with an exception in cases the thread is not in a +cancellable state. + +This method works a bit like throwing an exception that cannot be caught +- specifically, it will clean up the thread from within itself, so +all cleanup handlers (e.g. C blocks) are run with full thread +context and can block if they wish. The downside is that there is no +guarantee that the thread can be cancelled when you call this method, and +therefore, it might fail. It is also considerably slower than C or +C. + +A thread is in a safe-cancellable state if it either hasn't been run yet, +or it has no C context attached and is inside an SLF function. + +The latter two basically mean that the thread isn't currently inside a +perl callback called from some C function (usually via some XS modules) +and isn't currently executing inside some C function itself (via Coro's XS +API). + +This call returns true when it could cancel the thread, or croaks with an +error otherwise (i.e. it either returns true or doesn't return at all). + +Why the weird interface? Well, there are two common models on how and +when to cancel things. In the first, you have the expectation that your +coro thread can be cancelled when you want to cancel it - if the thread +isn't cancellable, this would be a bug somewhere, so C<< ->safe_cancel >> +croaks to notify of the bug. + +In the second model you sometimes want to ask nicely to cancel a thread, +but if it's not a good time, well, then don't cancel. This can be done +relatively easy like this: - if ($current == $self) { - terminate @_; - } else { - $self->{_status} = [@_]; - Coro::State::cancel $self; + if (! eval { $coro->safe_cancel }) { + warn "unable to cancel thread: $@"; } -} + +However, what you never should do is first try to cancel "safely" and +if that fails, cancel the "hard" way with C<< ->cancel >>. That makes +no sense: either you rely on being able to execute cleanup code in your +thread context, or you don't. If you do, then C<< ->safe_cancel >> is the +only way, and if you don't, then C<< ->cancel >> is always faster and more +direct. =item $coro->schedule_to @@ -546,17 +833,18 @@ Coro will check for the exception each time a schedule-like-function returns, i.e. after each C, C, C<< Coro::Semaphore->down ->>, C<< Coro::Handle->readable >> and so on. Most of these functions -detect this case and return early in case an exception is pending. +>>, C<< Coro::Handle->readable >> and so on. Most of those functions (all +that are part of Coro itself) detect this case and return early in case an +exception is pending. The exception object will be thrown "as is" with the specified scalar in C<$@>, i.e. if it is a string, no line number or newline will be appended (unlike with C). -This can be used as a softer means than C to ask a coro to -end itself, although there is no guarantee that the exception will lead to -termination, and if the exception isn't caught it might well end the whole -program. +This can be used as a softer means than either C or Cto ask a coro to end itself, although there is no guarantee that the +exception will lead to termination, and if the exception isn't caught it +might well end the whole program. You might also think of C as being the moral equivalent of Cing a coro with a signal (in this case, a scalar). @@ -565,12 +853,12 @@ Wait until the coro terminates and return any values given to the C or C functions. C can be called concurrently -from multiple coro, and all will be resumed and given the status +from multiple threads, and all will be resumed and given the status return once the C<$coro> terminates. =cut -sub join { +sub xjoin { my $self = shift; unless ($self->{_status}) { @@ -584,18 +872,22 @@ &schedule while $current; } - wantarray ? @{$self->{_status}} : $self->{_status}[0]; + wantarray ? @{$self->{_status}} : $self->{_status}[0] } =item $coro->on_destroy (\&cb) -Registers a callback that is called when this coro gets destroyed, -but before it is joined. The callback gets passed the terminate arguments, -if any, and I die, under any circumstances. +Registers a callback that is called when this coro thread gets destroyed, +that is, after it's resources have been freed but before it is joined. The +callback gets passed the terminate/cancel arguments, if any, and I die, under any circumstances. + +There can be any number of C callbacks per coro, and there is +no way currently to remove a callback once added. =cut -sub on_destroy { +sub xon_destroy { my ($self, $cb) = @_; push @{ $self->{_on_destroy} }, $cb; @@ -604,8 +896,8 @@ =item $oldprio = $coro->prio ($newprio) Sets (or gets, if the argument is missing) the priority of the -coro. Higher priority coro get run before lower priority -coro. Priorities are small signed integers (currently -4 .. +3), +coro thread. Higher priority coro get run before lower priority +coros. Priorities are small signed integers (currently -4 .. +3), that you can refer to using PRIO_xxx constants (use the import tag :prio to get then): @@ -615,27 +907,38 @@ # set priority to HIGH current->prio (PRIO_HIGH); -The idle coro ($Coro::idle) always has a lower priority than any +The idle coro thread ($Coro::idle) always has a lower priority than any existing coro. Changing the priority of the current coro will take effect immediately, -but changing the priority of coro in the ready queue (but not -running) will only take effect after the next schedule (of that -coro). This is a bug that will be fixed in some future version. +but changing the priority of a coro in the ready queue (but not running) +will only take effect after the next schedule (of that coro). This is a +bug that will be fixed in some future version. =item $newprio = $coro->nice ($change) Similar to C, but subtract the given value from the priority (i.e. -higher values mean lower priority, just as in unix). +higher values mean lower priority, just as in UNIX's nice command). =item $olddesc = $coro->desc ($newdesc) Sets (or gets in case the argument is missing) the description for this -coro. This is just a free-form string you can associate with a +coro thread. This is just a free-form string you can associate with a coro. This method simply sets the C<< $coro->{desc} >> member to the given -string. You can modify this member directly if you wish. +string. You can modify this member directly if you wish, and in fact, this +is often preferred to indicate major processing states that cna then be +seen for example in a L session: + + sub my_long_function { + local $Coro::current->{desc} = "now in my_long_function"; + ... + $Coro::current->{desc} = "my_long_function: phase 1"; + ... + $Coro::current->{desc} = "my_long_function: phase 2"; + ... + } =cut @@ -682,11 +985,16 @@ original code ref will be called (with parameters) from within another coro. -The reason this function exists is that many event libraries (such as the -venerable L module) are not thread-safe (a weaker form +The reason this function exists is that many event libraries (such as +the venerable L module) are not thread-safe (a weaker form of reentrancy). This means you must not block within event callbacks, otherwise you might suffer from crashes or worse. The only event library -currently known that is safe to use without C is L. +currently known that is safe to use without C is L (but +you might still run into deadlocks if all event loops are blocked). + +Coro will try to catch you when you block in the event loop +("FATAL:$Coro::IDLE blocked itself"), but this is just best effort and +only works when you do not run your own event loop. This function allows your callbacks to block by executing them in another coro where it is safe to block. One example where blocking is handy @@ -738,7 +1046,7 @@ } } -=item $cb = Coro::rouse_cb +=item $cb = rouse_cb Create and return a "rouse callback". That's a code reference that, when called, will remember a copy of its arguments and notify the owner @@ -746,14 +1054,16 @@ See the next function. -=item @args = Coro::rouse_wait [$cb] +=item @args = rouse_wait [$cb] Wait for the specified rouse callback (or the last one that was created in this coro). As soon as the callback is invoked (or when the callback was invoked before C), it will return the arguments originally passed to -the rouse callback. +the rouse callback. In scalar context, that means you get the I +argument, just as if C had a C +statement at the end. See the section B for an actual usage example. @@ -761,6 +1071,20 @@ =cut +for my $module (qw(Channel RWLock Semaphore SemaphoreSet Signal Specific)) { + my $old = defined &{"Coro::$module\::new"} && \&{"Coro::$module\::new"}; + + *{"Coro::$module\::new"} = sub { + require "Coro/$module.pm"; + + # some modules have their new predefined in State.xs, some don't + *{"Coro::$module\::new"} = $old + if $old; + + goto &{"Coro::$module\::new"}; + }; +} + 1; =head1 HOW TO WAIT FOR A CALLBACK @@ -849,10 +1173,14 @@ the windows process emulation enabled under unix roughly halves perl performance, even when not used. +Attempts to use threads created in another emulated process will crash +("cleanly", with a null pointer exception). + =item coro switching is not signal safe -You must not switch to another coro from within a signal handler -(only relevant with %SIG - most event libraries provide safe signals). +You must not switch to another coro from within a signal handler (only +relevant with %SIG - most event libraries provide safe signals), I +you are sure you are not interrupting a Coro function. That means you I call any function that might "block" the current coro - C, C C<< Coro::Semaphore->down >> or @@ -862,6 +1190,67 @@ =back +=head1 WINDOWS PROCESS EMULATION + +A great many people seem to be confused about ithreads (for example, Chip +Salzenberg called me unintelligent, incapable, stupid and gullible, +while in the same mail making rather confused statements about perl +ithreads (for example, that memory or files would be shared), showing his +lack of understanding of this area - if it is hard to understand for Chip, +it is probably not obvious to everybody). + +What follows is an ultra-condensed version of my talk about threads in +scripting languages given on the perl workshop 2009: + +The so-called "ithreads" were originally implemented for two reasons: +first, to (badly) emulate unix processes on native win32 perls, and +secondly, to replace the older, real thread model ("5.005-threads"). + +It does that by using threads instead of OS processes. The difference +between processes and threads is that threads share memory (and other +state, such as files) between threads within a single process, while +processes do not share anything (at least not semantically). That +means that modifications done by one thread are seen by others, while +modifications by one process are not seen by other processes. + +The "ithreads" work exactly like that: when creating a new ithreads +process, all state is copied (memory is copied physically, files and code +is copied logically). Afterwards, it isolates all modifications. On UNIX, +the same behaviour can be achieved by using operating system processes, +except that UNIX typically uses hardware built into the system to do this +efficiently, while the windows process emulation emulates this hardware in +software (rather efficiently, but of course it is still much slower than +dedicated hardware). + +As mentioned before, loading code, modifying code, modifying data +structures and so on is only visible in the ithreads process doing the +modification, not in other ithread processes within the same OS process. + +This is why "ithreads" do not implement threads for perl at all, only +processes. What makes it so bad is that on non-windows platforms, you can +actually take advantage of custom hardware for this purpose (as evidenced +by the forks module, which gives you the (i-) threads API, just much +faster). + +Sharing data is in the i-threads model is done by transfering data +structures between threads using copying semantics, which is very slow - +shared data simply does not exist. Benchmarks using i-threads which are +communication-intensive show extremely bad behaviour with i-threads (in +fact, so bad that Coro, which cannot take direct advantage of multiple +CPUs, is often orders of magnitude faster because it shares data using +real threads, refer to my talk for details). + +As summary, i-threads *use* threads to implement processes, while +the compatible forks module *uses* processes to emulate, uhm, +processes. I-threads slow down every perl program when enabled, and +outside of windows, serve no (or little) practical purpose, but +disadvantages every single-threaded Perl program. + +This is the reason that I try to avoid the name "ithreads", as it is +misleading as it implies that it implements some kind of thread model for +perl, and prefer the name "windows process emulation", which describes the +actual use and behaviour of it much better. + =head1 SEE ALSO Event-Loop integration: L, L, L.