postTo() sets up two-way communication: the caller asks to run work on some
other WorkQueue, expecting an eventual callback on the originating WorkQueue.
That permits us to transport any exception thrown by the work callable back to
rethrow on the originating WorkQueue.
In addition to the name making the blocking explicit, we changed the
signature: instead of specifying a target WorkQueue on which to run,
waitForResult() runs the passed callable on its own WorkQueue.
Why is that? Because, unlike postTo(), we do not require a handshake between
two different WorkQueues. postTo() allows running arbitrary callback code,
setting variables or whatever, on the originating WorkQueue (presumably on the
originating thread). waitForResult() synchronizes using Promise/Future, which
are explicitly designed for cross-thread communication. We need not call
set_value() on the originating thread, so we don't need a postTo() callback
lambda.
The idea is that you can call runOn(target, callable) from a (non-default)
coroutine and block that coroutine until the result becomes available.
As a safety check, we forbid calling runOn() from a thread's default
coroutine, assuming that a given thread's default coroutine is the one
servicing the relevant WorkQueue.
ThreadPool bundles a WorkQueue with the specified number of worker threads to
service it. Each ThreadPool has a name that can be used to locate its
WorkQueue.
Each worker thread calls WorkQueue::runUntilClose().
ThreadPool listens on the "LLApp" LLEventPump for shutdown notification. On
receiving that, it closes its WorkQueue and then join()s each of its worker
threads for orderly shutdown.
Add a settings.xml entry "ThreadPoolSizes", the first LLSD-valued settings
entry to expect a map: pool name->size. The expectation is that usually code
instantiating a particular ThreadPool will have a default size in mind, but it
should check "ThreadPoolSizes" for a user override.
Make idle_startup()'s STATE_SEED_CAP_GRANTED state instantiate a "General"
ThreadPool. This is function-static for lazy initialization.
Eliminate LLMainLoopRepeater, which is completely unreferenced. Any potential
future use cases are better addressed by posting to the main loop's WorkQueue.
Eliminate llappviewer.cpp's private LLDeferredTaskList class, which
implemented LLAppViewer::addOnIdleCallback(). Make addOnIdleCallback() post
work to the main loop's WorkQueue instead.
Make LLAppViewer::idle() call LL::WorkQueue::runFor() to dequeue and run some
or all of the pending responses from worker threads.
Add a MainWorkTime setting to specify the time slice the main loop may devote
each frame to servicing such responses.
A typical WorkQueue has a string name, which can be used to find it to post
work to it. "Work" is a nullary callable.
WorkQueue is a multi-producer, multi-consumer thread-safe queue: multiple
threads can service the WorkQueue, multiple threads can post work to it.
Work can be scheduled in the future by submitting with a timestamp. In
addition, a given work item can be scheduled to run on a recurring basis.
A requesting thread servicing a WorkQueue of its own, such as the viewer's
main thread, can submit work to another WorkQueue along with a callback to be
passed the result (of arbitrary type) of the first work item. The callback is
posted to the originating WorkQueue, permitting safe data exchange between
participating threads.
Methods are provided for different kinds of servicing threads. runUntilClose()
is useful for a simple worker thread. runFor(duration) devotes no more than a
specified time slice to that WorkQueue, e.g. for use by the main thread.
Its previous behavior, returning a const reference without locking, was wrong:
it could return a reference to an object in an inconsistent state if it was
concurrently being modified on another thread.
Locking the mutex and returning a copy by value is the correct behavior.
It feels wrong to return a dumb LLInstanceTracker subclass* from getInstance()
when we use std::shared_ptr and std::weak_ptr internally. But tweak consumers
to use 'auto' or LLInstanceTracker::ptr_t in case we later revisit this
decision.
We did add a couple get() calls where it's important to obtain a dumb pointer.
ThreadSafeSchedule::tryPopUntil() (and therefore tryPopFor()) was simply
delegating to LLThreadSafeQueue::tryPopUntil(), with an adjusted timeout since
we want to wake up as soon as the head item, if any, becomes ready. But then
we have to loop back to retry the pop to actually deal with that head item.
In addition, ThreadSafeSchedule::popWithTime() was spinning rather than
properly blocking on a timed condition variable. Fixed.
ThreadSafeSchedule orders its items by timestamp, which can be passed either
implicitly or explicitly. The timestamp specifies earliest delivery time: an
item cannot be popped until that time.
Add initial tests.
Tweak the LLThreadSafeQueue base class to support ThreadSafeSchedule:
introduce virtual canPop() method to report whether the current head item is
available to pop. The base class unconditionally says yes, ThreadSafeSchedule
says it depends on whether its timestamp is still in the future.
This replaces the protected pop_() overload accepting a predicate. Rather than
explicitly passing a predicate through a couple levels of function call, use
canPop() at the level it matters. Runtime behavior that varies depending on
an object's leaf class is what virtual functions were invented for.
Give pop_() a three-state enum return so pop() can distinguish between "closed
and empty" (throws exception) versus "closed, not yet drained because we're
not yet ready to pop the head item" (waits).
Also break out protected tryPopUntil_() method, the body logic of
tryPopUntil(). The public method locks the data structure, the protected
method requires that its caller has already done so.
Add chrono.h with a more full-featured LL::time_point_cast() function than the
one found in <chrono>, which only converts between time_point durations, not
between time_points based on different clocks.
Add LL::PriorityQueueAdapter, a wrapper for std::priority_queue to make its
API more closely resemble std::queue for drop-in use as LLThreadSafeQueue's
underlying QueueT container.
Support move-only element types.
Factor out some implementation redundancy: wrap actual push semantics as
push_(), actual pop semantics as pop_(). push(), tryPush() and tryPushUntil()
now call push_(); pop(), tryPop() and tryPopUntil() now call pop_().
Break out tryLock() and tryLockUntil() methods that, if they can lock, run the
passed callable. Then tryPush(), tryPushUntil(), tryPop() and tryPopUntil()
pass lambdas containing the meat of the original method body to tryLock() or
tryLockUntil(), as appropriate.
These functions allow prepending or removing an item at the left end of an
arbitrary tuple -- for instance, to add a sequence key to a caller's data,
then remove it again when delivering the original tuple.
First, parameterize LLThreadSafeQueue's queue type. This allows us to
substitute (e.g.) a std::priority_queue for a particular instance.
Use std::queue for the default queue type, changing the operations invoked on
the queue type from std::deque methods to std::queue methods.
Rename published methods from (e.g.) pushFront() and popBack() to simple
push() and pop(), retaining legacy names as aliases. Not only are the overt
Front and Back unnecessary; they're the opposite of how std::queue uses
std::deque or std::list, so they only confuse the reader.
Break out tryPushUntil() method. We already use that logic internally to
tryPushFor(), so it's just as easy to publish it as its own entry point.
Add tryPopFor() and tryPopUntil() to allow limiting the time we'll wait for a
queue item to become available.
The generic build.sh script honors these environment variables to specify
parameters to the autobuild configure and autobuild build steps, respectively.
Support them in the viewer-specific build.sh too.
In the generic build.sh, autobuild_configure_parameters allows specifying
command-line switches either for autobuild or, following --, for the
underlying tool (in our case, CMake). In order to support that variable the
same way here, we insert -- (as before) when autobuild_configure_parameters is
unset or empty, since the rest of the switches *we* specify are for CMake.
That means that, as with the generic build.sh, a non-empty
autobuild_configure_parameters override must precede any switches intended for
CMake with the -- separator.
Nat Goodspeed