Specifically, llcoro::suspendUntilTimeout() is definitely called concurrently
by multiple coroutines. New code that instantiates a local LLEventStream must
allow the name to be tweaked for uniqueness.
Specifically, the shutdown crash reported in SL-13512 was due to
LLExperienceCache::idleCoro() looping on suspendUntilTimeout(), failing to
notice in its slumbers that the viewer was shutting down around it.
Make suspendUntilTimeout() internally call suspendUntilEventOnWithTimeout(),
which already listens for "LLApp" state-change events and throws Stopping when
LLApp enters its shutdown sequence.
The new LLCoros::Stop exception is intended to terminate long-lived coroutines
-- not interrupt mainstream shutdown processing. Only throw it on an
explicitly-launched coroutine.
Make LLCoros::getName() (used by the above test) static. As with other LLCoros
methods, it might be called after the LLCoros LLSingleton instance has been
deleted. Requiring the caller to call instance() implies a possible need to
also call wasDeleted(). Encapsulate that nuance into a static method instead.
By the time "LLApp" listeners are notified that the app is quitting, the
mainloop is no longer running. Even though those listeners do things like
close work queues and inject exceptions into pending promises, any coroutines
waiting on those resources must regain control before they can notice and shut
down properly. Add a final "LLApp" listener that resumes ready coroutines a
few more times.
Make sure every other "LLApp" listener is positioned before that new one.
Add LLCoros::TempStatus instances around known suspension points so
printActiveCoroutines() can report what each suspended coroutine is waiting
for.
Similarly, sprinkle checkStop() calls at known suspension points.
Make LLApp::setStatus() post an event to a new LLEventPump "LLApp" with a
string corresponding to the status value being set, but only until
~LLEventPumps() -- since setStatus() also gets called very late in the
application's lifetime.
Make postAndSuspendSetup() (used by postAndSuspend(), suspendUntilEventOn(),
postAndSuspendWithTimeout(), suspendUntilEventOnWithTimeout()) add a listener
on the new "LLApp" LLEventPump that pushes the new LLCoros::Stopping exception
to the coroutine waiting on the LLCoros::Promise. Make it return the new
LLBoundListener along with the previous one.
Accordingly, make postAndSuspend() and postAndSuspendWithTimeout() store the
new LLBoundListener returned by postAndSuspendSetup() in a LLTempBoundListener
(as with the previous one) so it will automatically disconnect once the wait
is over.
Make each LLCoprocedurePool instance listen on "LLApp" with a listener that
closes the queue on which new work items are dispatched. Closing the queue
causes the waiting dispatch coroutine to terminate. Store the connection in an
LLTempBoundListener on the LLCoprocedurePool so it will disconnect
automatically on destruction.
Refactor the loop in coprocedureInvokerCoro() to instantiate TempStatus around
the suspending call.
Change a couple spammy LL_INFOS() calls to LL_DEBUGS(). Give all logging calls
in that module a "CoProcMgr" tag to make it straightforward to re-enable the
LL_DEBUGS() calls as desired.
Although this is legal, apparently there was a bug in the C++11 standard (to
which gcc 4.8 conforms) that was subsequently fixed in the standard (and thus
in gcc 4.9). Thanks Henri Beauchamp.
passed to postAndSuspendsetup().
The requestPump is optional, and the function varies its behavior depending on
whether that parameter is empty or meaningful. But it unconditionally uses the
replyPump. Passing an empty LLEventPumpOrPumpName caused mysterious crashes.
Add llassert_always_msg() to make the coding error explicit in such a case.
Also streamline access to meaningful requestPump and replyPump by temporarily
caching the bound LLEventPump reference.
Actually the fix is in postAndSuspendSetup(), which affects postAndSuspend(),
postAndSuspendWithTimeout(), suspendUntilEventOnWithTimeout() and
suspendUntilEventOn().
By "overflow case" we mean the special circumstance in which:
* the LLEventPump in question is an LLEventMailDrop, meaning its listeners
eventually expect to see every post()ed value
* one of the listeners is supposed to consume those values (has called
LLCoros::set_consuming(true))
* post() is called more than once before that listener is resumed.
The magic of postAndSuspend() (et al.) is a temporary LLCoros::Promise. The
waiting coroutine calls get() on the corresponding Future, causing it to
suspend (as promised) until the Promise is fulfilled.
With the Boost.Fiber implementation of coroutines, fulfilling the Promise
doesn't immediately resume the suspended coroutine -- it merely marks it ready
to resume, next time the scheduler gets control.
A second post() call before the suspended coroutine is resumed results in a
second call to Promise::set_value(). But Promise is a one-shot entity. This
results in a promise_already_satisfied exception. Because a second post() call
during that time window is perfectly reasonable, we catch that exception and
carry on.
The tricky part is: when that exception is thrown, what should the listener
return? Previously we were returning the listener's current consuming setting,
just as when the set_value() call succeeds.
But when the LLEventPump is an LLEventMailDrop, and the listener's consuming
flag is true, that told LLEventMailDrop::post() that the value got through,
and that it needn't bother to save it in its history queue. The net effect was
to discard the value.
Instead, return the listener's consuming flag only when Promise::set_value()
succeeds. When it throws promise_already_satisfied, unconditionally return
false. That directs LLEventMailDrop::post() to enqueue the undelivered value
so that the *next* suspendUntilEventOn() call can pick it up.
Longtime fans will remember that the "dcoroutine" library is a Google Summer
of Code project by Giovanni P. Deretta. He originally called it
"Boost.Coroutine," and we originally added it to our 3p-boost autobuild
package as such. But when the official Boost.Coroutine library came along
(with a very different API), and we still needed the API of the GSoC project,
we renamed the unofficial one "dcoroutine" to allow coexistence.
The "dcoroutine" library had an internal low-level API more or less analogous
to Boost.Context. We later introduced an implementation of that internal API
based on Boost.Context, a step towards eliminating the GSoC code in favor of
official, supported Boost code.
However, recent versions of Boost.Context no longer support the API on which
we built the shim for "dcoroutine." We started down the path of reimplementing
that shim using the current Boost.Context API -- then realized that it's time
to bite the bullet and replace the "dcoroutine" API with the Boost.Fiber API,
which we've been itching to do for literally years now.
Naturally, most of the heavy lifting is in llcoros.{h,cpp} and
lleventcoro.{h,cpp} -- which is good: the LLCoros layer abstracts away most of
the differences between "dcoroutine" and Boost.Fiber.
The one feature Boost.Fiber does not provide is the ability to forcibly
terminate some other fiber. Accordingly, disable LLCoros::kill() and
LLCoprocedureManager::shutdown(). The only known shutdown() call was in
LLCoprocedurePool's destructor.
We also took the opportunity to remove postAndSuspend2() and its associated
machinery: FutureListener2, LLErrorEvent, errorException(), errorLog(),
LLCoroEventPumps. All that dual-LLEventPump stuff was introduced at a time
when the Responder pattern was king, and we assumed we'd want to listen on one
LLEventPump with the success handler and on another with the error handler. We
have never actually used that in practice. Remove associated tests, of course.
There is one other semantic difference that necessitates patching a number of
tests: with "dcoroutine," fulfilling a future IMMEDIATELY resumes the waiting
coroutine. With Boost.Fiber, fulfilling a future merely marks the fiber as
ready to resume next time the scheduler gets around to it. To observe the test
side effects, we've inserted a number of llcoro::suspend() calls -- also in
the main loop.
For a long time we retained a single unit test exercising the raw "dcoroutine"
API. Remove that.
Eliminate llcoro_get_id.{h,cpp}, which provided llcoro::get_id(), which was a
hack to emulate fiber-local variables. Since Boost.Fiber has an actual API for
that, remove the hack.
In fact, use (new alias) LLCoros::local_ptr for LLSingleton's dependency
tracking in place of llcoro::get_id().
In CMake land, replace BOOST_COROUTINE_LIBRARY with BOOST_FIBER_LIBRARY. We
don't actually use the Boost.Coroutine for anything (though there exist
plausible use cases).
We include both const and non-const overloads. The latter returns LLSD&, so
you can assign to the located element.
In fact we already implemented the non-const logic in a less public form as
storeToLLSDPath() in lleventcoro.cpp. Reimplement the latter to use the new
llsd::drill() function.
This also introduces LLContinueError for exceptions which should interrupt
some part of viewer processing (e.g. the current coroutine) but should attempt
to let the viewer session proceed.
Derive all existing viewer exception classes from LLException rather than from
std::runtime_error or std::logic_error.
Use BOOST_THROW_EXCEPTION() rather than plain 'throw' to enrich the thrown
exception with source file, line number and containing function.
The original implementation of set_consuming() involved a bool* pointing to a
local bool in VoidListener::operator()()'s stack frame. postAndSuspend() would
set that bool (through the pointer) as soon as it returned from suspension.
The trouble with that is that LLEventMailDrop potentially calls its new
listener (fulfilling the future) immediately in the listen_impl() override --
in other words, way up at the top of postAndSuspend(), well before the code
that sets the relevant bool.
Instead, make the adapter formerly known as VoidListener bind the coroutine's
get_consuming() value at adapter construction time (before listening on the
LLEventPump), so that its operator()() has the coroutine's correct
get_consuming() value to return. Eliminating the bool* makes the code both
simpler AND more correct!
This change makes that adapter very specific to coroutine usage. Rename it
FutureListener and migrate it from lleventcoros.h into the .cpp file. Nobody
else was using it anyway.
Make corresponding changes to postAndSuspend2() and its WaitForEventOnHelper
class -- whose name no longer corresponds to the function as it used to.
Rename that one FutureListener2. The new FutureListener functionality, common
to both these adapters, makes it useful to derive FutureListener2 from
FutureListener.
Introduce llmake(), a generic function to deduce template type arguments from
function parameter types. This allows us to remove the voidlistener() and
wfeoh() helper functions.
Hiding VoidListener broke one of the lleventcoro_test.cpp tests. But that test
was sort of a lame recap of an earlier implementation of postAndSuspend(),
based on LLEventPump events. Recast that test to illustrate how to use a
coroutine future to suspend a coroutine for something other than an LLEventPump.
But that rubbed my nose in the fact that we MUST wrap future's context
switching with proper management of the current coroutine. Introduce
LLCoros::Future<T>, which wraps boost::dcoroutines::future<T>.
Use LLCoros::Future<T> in postAndSuspend() and postAndSuspend2().
Our first cut at tracking the "current" coroutine simply reset the pointer to
NULL every time we context-switched away. But that strategy doesn't handle the
case of coroutine A launching coroutine B.
Introduce LLCoros::CoroData to track, among other things, the previous value
of the current-coroutine pointer each time we switch into a coroutine. Restore
THAT value when we switch back out.
To date, the coroutine helper functions in lleventcoro.h have been in the
global namespace. Migrate them into llcoro namespace, and fix references.
Specifically, LLVoidListener => llcoro::VoidListener, and voidlistener(),
postAndWait(), both waitForEventOn(), postAndWait2(), errorException() and
errorLog() have been moved into llcoro.
Also migrate new LLCoros::get_self() and Suspending to llcoro:: namespace.
While at it, I realized that -- having converted several lleventcoro.h
functions from templates (for arbitrary 'self' parameter type) to ordinary
functions, having moved them from lleventcoro.h to lleventcoro.cpp, we can now
migrate their helpers from lleventcoro.h to lleventcoro.cpp as well. This
eliminates the need for the LLEventDetail namespace; the relevant helpers are
now in an anonymous namespace in the .cpp file: listenerNameForCoro(),
storeToLLSDPath(), WaitForEventOnHelper and wfeoh().
instances. LLCoros::launch() intends to address three issues:
- ownership of coroutine instance
- cleanup of coroutine instance when it terminates
- central place to twiddle MSVC optimizations to bypass DEV-32777 crash.
Initially coded on Mac; will address the third bullet on Windows.
Adapt listenerNameForCoro() to consult LLCoros::getName() if applicable.
Change LLLogin::Impl::connect() to use LLCoros::launch().
LLCoros::getName() relies on patch to boost::coroutines::coroutine::self to
introduce get_id().
Andrey Lihatskiy