Sunday, June 16, 2013

App reference, Perl, part 1

The App API has two parts to it: The first part keeps track of all the App instances in the program, and allows to list and find them. The second part is the manipulations on a particular instance.

The global part of the API is:

@apps = Triceps::App::listApps();

Returns the array of name-value pairs, values containing the App references, listing all the Apps in the program (more exactly, in this Triceps library, you you compile multiple Triceps libraries together by renaming them, each of them will have its own Apps list). The returned array can be placed into a hash.

$app = Triceps::App::find($name);

Find an App by name. If an App with such a name does not exist, it will confess.

$app = Triceps::App:make($name);

Create a new App, give it the specified name, and put it on the list. The names must not be duplicate with the other existing Apps, or the method will confess.


Drop the app, by reference or by name, from the list. The App is as usual reference-counted and will exist while there are references to it. The global list provides one of these references, so an App is guaranteed to exist while it's still on the list. When dropped, it will still be accessible through the existing references, but obviously will not be listed any more and could not be found by name.

Moreover, a new App with the same name can be added to the list. Because of this, dropping an App by name requires some care in case if there could be a new App created again with the same name: it creates a potential for a race, and you might end up dropping the new App instead of the old one. Of course, if it's the same thread that drops the old one and creates the new one, then there is no race. Dropping an application by name that doesn't exist at the moment is an error and will confess.

Dropping the App by reference theoretically allows to avoid the race: a specific object gets dropped, and if it already has been dropped, the call has no effect. Theoretically. However in practice Perl has a limitation of passing the object values between threads, and thus whenever each thread starts, first thing it does is finding its App by name. It's a very similar kind of race and is absolutely unavoidable except by making sure that all the App's threads have exited and joined (i.e. harvesting them). So make sure to complete the App's thread harvesting before dropping the App in the harvester thread, and by then it doesn't matter a whole lot if the App is dropped by name or by reference.

Now the second part of the API, working with an App instance.

Many (but not all) of the App methods allow to specify the App either by reference or by name, and they automatically sort it out, doing the internal look-up by name if necessary. So the same method could be used as either of:

Triceps::App::method($app, ...);
Triceps::App::method($appName, ...);

Obviously, you can not use the "->" syntax with a name, and obviously if the name is not in the app list, the method will confess. Below I'll show the calls that allow the dual formats as Triceps::App::method($appOrName, ...) but keep in mind that you can use the "->" form of them too with a reference.

$app = Triceps::App::resolve($appOrName);
Do just the automatically-sorting-out part: gets a reference or name and returns a reference either way. A newly created reference is returned in either case (not the argument reference). You can use this resolver before the methods that accept only a reference.

$result = $app->same($app2);

Check if two references are for the same App object. Here they both must be references and not names.

$name = $app->getName();

Get the name of the App, from a reference.

Triceps::App::declareTriead($appOrName, $trieadName);

Declare a Triead (Triceps thread) in advance. Any attempts to look up the nexuses in that thread will then wait for the thread to become ready. (Attempts to look up in an undeclared and undefined thread are errors). This is necessary to prevent a race at the thread creation time.  For the most part, the method Triead::start() just does the right thing by calling this method automatically and you don't need the use it manually, except in some very special circumstances.

@trieads = Triceps::App::getTrieads($appOrName);

Get the list of currently defined Trieads, as name-value pairs. Keep in mind that the other threads may be modifying the list of Trieads, so if you do this call multiple times, you may get different results. However the Trieads are returned as references, so they are guaranteed to stay alive and readable even if they get removed from the App, or even if the App gets dropped.


Run the harvester in the current thread. The harvester gets notifications from the threads when they are about to exit, and joins them.  After all the threads have been joined, it automatically drops the App, and returns.
The harvesting is an absolutely necessary part of the App life cycle, however in most of the usage patterns (such as with Triead::startHere or App::build) the harvester is called implicitly from the wrapping library functions, so you don't need to care about it.

Note also that if you're running the harvester manually, you must call it only after the first thread has been defined or at least declared. Otherwise it will find no threads in the App, consider it dead and immediately drop it.

If the App was aborted, the harvester will normally confess after if had joined all the threads and disposed of the App, unless the option die_on_abort (see below) has been set to 0. This propagates the error to the caller. However there is a catch: if some of the threads don't react properly by exiting on an abort indication, the program will be stuck and you will see no error message until these threads get unstuck, possibly forever.


 die_on_abort => 1/0

(default: 1) If the App was aborted, the harvester will normally confess after if had joined all the threads and disposed of the App.  Setting this option to 0 will make the aborts silently ignored. This option does not affect the errors in joining the threads: if any of those are detected, harvester will still confess after it had disposed of the app.

$dead = $app->harvestOnce();

Do one run of the harvesting.  Joins all the threads that have exited since its last call. If no threads have exited since then, returns immediately. Returns 1 if all the threads have exited (and thus the App is dead), 0 otherwise.  If a thread join fails, immediately confesses (if multiple threads were ready for joining, the ones queued after the failed one won't be joined, call harvestOnce() once more to join them).


Wait for at least one thread to become ready for harvesting. If the App is already dead (i.e. all its threads have exited), returns immediately.

These two methods allow to write the custom harvesters if you're not happy with the default one. The basic harvester logic can be written as:

do {
} while(!$app->harvestOnce());

However the real harvester also does some smarter things around the error handling. You can look it up in the source code in cpp/app/App.cpp.

$res = Triceps::App::isDead($appOrName);

Returns 1 if the App is dead (i.e. has no more live threads), otherwise 0.  Calling this method with a name for the argument is probably a bad idea, since normally the harvester will drop the App quickly after it becomes dead, and you may end up with this method confessing when it could not find the dropped App.

$res = Triceps::App::isShutdown($appOrName);

Returns 1  if the App has been requested to shut down, either normally or by being aborted.

$res = Triceps::App::isAborted($appOrName);

Returns 1 if the App has been aborted. The App may be aborted explicitly by calling the method abortBy(), or the thread wrapper logic automatically converts any unexpected deaths in the App's threads to the aborts. If any thread dies, this aborts the App, which in turn requests the other threads to die on their next thread-related call. Eventually the harvester collects them all and confesses, normally making the whole program die with an error.

($tname, $message) = Triceps::App::getAborted($appOrName);

Get the App abort information: name of the thread that caused the abort, and its error message.

Triceps::App::abortBy($appOrName, $tname, $msg);

Abort the application. The thread name and message will be remembered, and returned later by getAborted() or in the harvester. If abortBy() is called multiple times, only the first pair of thread name and message gets remembered. The reason is that on abort all the threads get interrupted in a fairly rough manner (all their ongoing and following calls to the threading API die), which typically causes them to call abortBy() as well, and there is no point in collecting these spurious messages.

The thread name here doesn't have to be the name of the actual thread that reports the issue. For example, if the thread creation as such fails (maybe because the OS limit on the thread count) that gets detected by the parent thread but reported in the name of the thread whose creation has failed. And in general you can pass just any string as the thread name, App itself doesn't care, just make it something that makes sense to you.

$res = Triceps::App::isDead($appOrName);

Returns 1 if the App is dead (i.e. it has no alive Trieads, all the defined and declared threads have exited). Right after the App is created, before the first Triead is created, the App is also considered dead, and becomes alive when the first Triead is declared or defined. If an App becomes dead later, when all the Trieads exit, it can still be brought back alive by creating more Trieads. But this is considered bad practice, and will cause a race with the harvester (if you want to do this, you have to make your own custom harvester).

$res = Triceps::App::isShutdown($appOrName);

Returns 1 if the App was requested to shut down. The Trieads might still run for some time, until they properly detect and process the shutdown, and exit. So this condition is not equivalent to Dead, althouh they are connected. If any new Trieads get started, they will be shut down right away and won't run.

To reiterate: if all the Trieads just exit by themselves, the App becomes dead but not shut down. You could still start more Trieads and bring the App back alive. If the App has been shut down, it won't become immediately dead, but it will send the shutdown indication to all the Trieads, and after all of them eventually exit, the App will become dead too. And after shutdown there is no way to bring the App back alive, since any new Trieads will be shut down right away (OK, there might be a short period until they detect the shutdown, so the App could spike as alive for a short time, but then will become dead again).


Will wait for the App to become dead and return after that. Make sure to not call waitDead() from any of App's Trieads: that would cause a deadlock.


Shut down the App. The shutdown state is sticky, so any repeated calls will have no effect. The call returns immediately and doesn't wait for the App to die. If you want to wait, call waitDead() afterwards. Make sure to not call waitDead() from a Triead: that would cause a deadlock.

Triceps::App::shutdownFragment($appOrName, $fragName);

Shut down a named fragment. This does not shut down the whole App, it just selectively shuts down the Trieads belonging to this fragment . See the explanation of the fragments in The fragment shutdown is not sticky: after a fragment has been shut down, it's possible to create another fragment with the same name. To avoid races, a fragment may be shut down only after all its Trieads are ready. So the caller Triead must call readyReady() before it calls shutdownFragment(). If any of the fragment's Trieads are not ready, the call will confess.

No comments:

Post a Comment