| Age | Commit message (Collapse) | Author |
|---|
|
|
|
In the previous design, the tick() method ran each task exactly once.
doPeriodically() was implemented by posting a functor that would, after
calling the specified callable, repost itself at (timestamp + interval).
The trouble with that design is that it required (interval > 0). A nonpositive
interval would result in looping over any timers with nonpositive repetition
intervals without ever returning from the tick() method.
To avoid that, doPeriodically() contained an llassert(interval > 0).
Unfortunately the viewer failed that constraint immediately at login, leading
to the suspicion that eliminating every such usage might require a protracted
search.
Lifting that restriction required a redesign. Now the priority queue stores a
callable returning bool, and the tick() method itself contains the logic to
repost a recurring task -- but defers doing so until after it stops looping
over ready tasks, ensuring that a task with a nonpositive interval will at
least wait until the following tick() call.
This simplifies not only doPeriodically(), but also doAtTime(). The previous
split of doAtTime() into doAtTime1() and doAtTime2() was only to accommodate
the needs of the Periodic functor class. Ditch Periodic.
Per feedback from NickyD, rename doAtTime() to scheduleAt(), which wraps its
passed nullary callable into a callable that unconditionally returns true (so
tick() will run it only once).
Rename the doAfterInterval() method to scheduleAfter(), which similarly wraps
its nullary callable. However, the legacy doAfterInterval() free function
remains. scheduleAfter() also loses its llassert(seconds > 0).
Rename the doPeriodically() method to scheduleRepeating(). However, the legacy
doPeriodically() free function remains.
Add internal scheduleAtRepeating(), whose role is to accept both a specific
timestamp and a repetition interval (which might be ignored, depending on the
callable). scheduleAtRepeating() now contains the real logic to add a task.
Rename getRemaining() to timeUntilCall(), hopefully resolving the question of
"remaining what?"
Expand the std::pair metadata stored in Timers's auxiliary unordered_map to a
Metadata struct containing the repetition interval plus two bools to mediate
deferred cancel() processing. Rename HandleMap to MetaMap, mHandles to mMeta.
Defend against the case when cancel(handle) is reached during the call to that
handle's callable. Meta::mRunning is set for the duration of that call. When
cancel() sees mRunning, instead of immediately deleting map entries, it sets
mCancel. Upon return from a task's callable, tick() notices mCancel and
behaves as if the callable returned true to stop the series of calls.
To guarantee that mRunning doesn't inadvertently remain set even in the case
of an exception, introduce local RAII class TempSet whose constructor accepts
a non-const variable reference and a desired value. The constructor captures
the current value and sets the desired value; the destructor restores the
previous value.
Defend against exception in a task's callable, and stop calling that task. Use
LOG_UNHANDLED_EXCEPTION() to report it.
|
|
|
|
If there is exactly one doPeriodically() entry on LLLater, the mQueue entry is
deleted before Periodic::operator()() is called. If Periodic's callable
returns false, it reinstates itself on mQueue for a future time. That's okay,
because the test for mQueue.empty(), and the consequent disconnect from
LLCallbackList, happens only after that Periodic call returns.
But at the moment Periodic reinstates itself on mQueue, mQueue happens to be
empty. That alerts doAtTime2() to register itself on LLCallbackList -- even
though at that moment it's already registered.
Semantically that's okay because assigning to the LLLater's
LLCallbackList::temp_handle_t should implicitly disconnect the previous
connection. But it's pointless to create a new connection and disconnect the
old one every time we call that lone Periodic.
Add a test for mLive.connected(); that way if we already have an
LLCallbackList connection, we retain it instead of replacing it with a new
one.
|
|
Instead of maintaining a whole separate unordered_map to look up target times,
make room in the HandleMap entry for the target time. There's still
circularity, but the split into doAtTime1() and doAtTime2() resolves it: since
doAtTime2() accepts the mHandles iterator created by doAtTime1(), doAtTime2()
can simply store the new mQueue handle_type into the appropriate slot.
Also sprinkle in a few more override keywords for consistency.
|
|
Some timer use cases need to know not only whether the timer is active, but
how much time remains before it (next) fires.
Introduce LLLater::mDoneTimes to track, for each handle, the timestamp at
which it's expected to fire. We can't just look up the target timestamp in
mQueue's func_at entry because there's no documented way to navigate from a
handle_type to a node iterator or pointer. Nor can we store it in mHandles
because of order dependency: we need the mDoneTimes iterator so we can bind it
into the Periodic functor for doPeriodically(), but we need the mQueue handle
to store in mHandles. If we could find the mQueue node from the new handle, we
could update the func_at entry after emplace() -- but if we could find the
mQueue node from a handle, we wouldn't need to store the target timestamp
separately anyway.
Split LLLater::doAtTime() into internal doAtTime1() and doAtTime2(): the first
creates an mDoneTimes entry and returns an iterator, the second finishes
creating new mQueue and mHandles entries based on that mDoneTimes entry.
This lets doPeriodically()'s Periodic bind the mDoneTimes iterator. Then
instead of continually incrementing an internal data member, it increments the
mDoneTimes entry to set the next upcoming timestamp. That lets getRemaining()
report the next upcoming timestamp rather than only the original one.
Add LLEventTimer::isRunning() and getRemaining(), forwarding to its LLLater
handle.
Fix various LLEventTimer subclass references to mEventTimer.stop(), etc.
Fix non-inline LLEventTimer subclass tick() overrides for bool, not BOOL.
Remove LLAppViewer::idle() call to LLEventTimer::updateClass(). Since
LLApp::stepFrame() already calls LLCallbackList::callFunctions(), assume we've
already handled that every tick.
|
|
The big idea is to reduce the number of per-tick callbacks asking, "Is it time
yet? Is it time yet?" We do that for LLEventTimer and LLEventTimeout.
LLLater presents doAtTime(LLDate), with doAfterInterval() and doPeriodically()
methods implemented using doAtTime(). All return handles. The free functions
doAfterInterval() and doPeriodically() now forward to the corresponding
LLLater methods.
LLLater also presents isRunning(handle) and cancel(handle).
LLLater borrows the tactic of LLEventTimer: while there's at least one running
timer, it registers an LLCallbackList tick() callback to service ready timers.
But instead of looping over all of them asking, "Are you ready?" it keeps them
in a priority queue ordered by desired timestamp, and only touches those whose
timestamp has been reached. Also, it honors a maximum time slice: once the
ready timers have run for longer than the limit, it defers processing other
ready timers to the next tick() call. The intent is to consume fewer cycles
per tick() call, both by the management machinery and the timers themselves.
Revamp LLCallbackList to accept C++ callables in addition to (classic C
function pointer, void*) pairs. Make addFunction() return a handle (different
than LLLater handles) that can be passed to a new deleteFunction() overload,
since std::function instances can't be compared for equality.
In fact, implement LLCallbackList using boost::signals2::signal, which provides
almost exactly what we want.
LLCallbackList continues to accept (function pointer, void*) pairs, but now
we store a lambda that calls the function pointer with that void*. It takes
less horsing around to create a C++ callable from a (function pointer, void*)
pair than the other way around. For containsFunction() and deleteFunction(),
such pairs are the keys for a lookup table whose values are handles.
Instead of having a static global LLCallbackList gIdleCallbacks, make
LLCallbackList an LLSingleton to guarantee initialization. For backwards
compatibility, gIdleCallbacks is now a macro for LLCallbackList::instance().
Move doOnIdleOneTime() and doOnIdleRepeating() functions to LLCallbackList
methods, but for backwards compatibility continue providing free functions.
Reimplement LLEventTimer using LLLater::doPeriodically(). One implication is
that LLEventTimer need no longer be derived from LLInstanceTracker, which we
used to iterate over all instances every tick. Give it start() and stop()
methods, since some subclasses (e.g. LLFlashTimer) used to call its member
LLTimer's start() and stop(). Remove updateClass():
LLCallbackList::callFunctions() now takes care of that.
Remove LLToastLifeTimer::start() and stop(), since LLEventTimer now provides
those. Remove getRemainingTimeF32(), since LLLater does not (yet) provide that
feature.
While at it, make LLEventTimer::tick() return bool instead of BOOL, and change
existing overrides.
Make LLApp::stepFrame() call LLCallbackList::callFunctions() instead of
LLEventTimer::updateClass().
We could have refactored LLEventTimer to use the mechanism now built into
LLLater, but frankly the LLEventTimer API is rather clumsy. You MUST derive a
subclass and override tick(), and you must instantiate your subclass on the
heap because, when your tick() override returns false, LLEventTimer deletes
its subclass instance. The LLLater API is simpler to use, and LLEventTimer is
much simplified by using it.
Merge lleventfilter.h's LLEventTimeoutBase into LLEventTimeout, and likewise
merge LLEventThrottleBase into LLEventThrottle. The separation was for
testability, but now that they're no longer based on LLTimer, it becomes
harder to use dummy time for testing. Temporarily skip tests based on
LLEventTimeoutBase and LLEventThrottleBase.
Instead of listening for LLEventPump("mainloop") ticks and using LLTimer,
LLEventTimeout now uses LLLater::doAfterInterval(). Instead of LLTimer and
LLEventTimeout, LLEventThrottle likewise now uses LLLater::doAfterInterval().
Recast a couple local LLEventTimeout pre-lambda callable classes with lambdas.
Dignify F64 with a new typedef LLDate::timestamp. LLDate heavily depends on
that as its base time representation, but there are those who question use of
floating-point for time. This is a step towards insulating us from any future
change.
|
|
|
|
Also change from boost::hof::is_invocable() to std::is_invocable().
|
|
|
|
# Conflicts:
# autobuild.xml
# indra/llcommon/llsys.cpp
|
|
|
|
# Conflicts:
# .github/workflows/build.yaml
|
|
Closing window correctly caused a significant amount of logout freezes
with no known reproes. Temporarily returning to old behavior were thread
was killes without closing window and will reenable in later maints to
hopefully get a scenario or at least more data of what is causing the
freeze.
|
|
Under debug LL_ERRS will show a message as well, but release won't show
anything and will quit silently so show a notification when applicable.
|
|
|
|
# Conflicts:
# indra/newview/llinventorygallery.cpp
|
|
|
|
|
|
Note that crash happened when setting LLProgressView::setMessage
|
|
|
|
|
|
|
|
|
|
|
|
# Conflicts:
# indra/newview/llchiclet.h
|
|
|
|
UIImgInvisibleUUID doesn't exist
Default normal for material is 'null'
|
|
1. After window closes viewer still takes some time to shut down, so
added splash screen to not confuse users (and to see if something gets
stuck)
2. Having two identical mWindowHandle caused confusion for me, so I
split them. It looks like there might have been issues with thread being
stuck because thread's handle wasn't cleaned up.
3. Made region clean mCacheMap immediately instead of spending time
making copies on shutdown
|
|
|
|
a preset...' option of the 'Preferences' floater
|
|
|
|
|
|
coroutines).
|
|
|
|
|
|
|
|
# Conflicts:
# indra/newview/fonts/DejaVu-license.txt
# indra/newview/fonts/DejaVuSans-Bold.ttf
# indra/newview/fonts/DejaVuSans-BoldOblique.ttf
# indra/newview/fonts/DejaVuSans-Oblique.ttf
# indra/newview/fonts/DejaVuSans.ttf
# indra/newview/fonts/DejaVuSansMono.ttf
|
|
# Conflicts:
# indra/newview/llspatialpartition.cpp
|
|
|
|
# Conflicts:
# indra/newview/llinventorygallery.cpp
# indra/newview/skins/default/xui/en/notifications.xml
|
|
|
|
# Conflicts:
# indra/llrender/llgl.cpp
# indra/llrender/llvertexbuffer.cpp
# indra/llui/llflatlistview.cpp
# indra/newview/lldrawpoolground.cpp
# indra/newview/llspatialpartition.cpp
# indra/newview/lltexturefetch.cpp
# indra/newview/llviewergenericmessage.cpp
# indra/newview/llviewertexture.cpp
# indra/newview/llvosky.cpp
# indra/newview/skins/default/xui/en/floater_preferences_graphics_advanced.xml
# indra/newview/skins/default/xui/en/floater_stats.xml
# indra/newview/skins/default/xui/en/floater_texture_fetch_debugger.xml
# indra/newview/skins/default/xui/en/notifications.xml
# indra/newview/skins/default/xui/en/panel_performance_preferences.xml
|
|
# Conflicts:
# indra/llcommon/CMakeLists.txt
# indra/newview/llspatialpartition.cpp
# indra/newview/llviewergenericmessage.cpp
# indra/newview/llvoavatar.cpp
|
|
We actively use event pumps's connections in threads, make sure nothing
modifies list of connections during reset.
And in case this doesn't fix the issue list affected pump before it
crashes to have a better idea of what is going on.
|
|
|
|
by making it thread_local.
|
|
|
|
|
|
|
