Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

120 lines
4.5 KiB
C++
Raw Permalink Normal View History

/* -*- Mode: C++; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 4 -*- */
/*
* This file is part of the LibreOffice project.
*
* This Source Code Form is subject to the terms of the Mozilla Public
* License, v. 2.0. If a copy of the MPL was not distributed with this
* file, You can obtain one at http://mozilla.org/MPL/2.0/.
*
* This file incorporates work covered by the following license notice:
*
* Licensed to the Apache Software Foundation (ASF) under one or more
* contributor license agreements. See the NOTICE file distributed
* with this work for additional information regarding copyright
* ownership. The ASF licenses this file to you under the Apache
* License, Version 2.0 (the "License"); you may not use this file
* except in compliance with the License. You may obtain a copy of
* the License at http://www.apache.org/licenses/LICENSE-2.0 .
*/
#ifndef INCLUDED_VCL_TASK_HXX
#define INCLUDED_VCL_TASK_HXX
#include <vcl/dllapi.h>
struct ImplSchedulerData;
enum class TaskPriority
{
HIGHEST, ///< These events should run very fast!
DEFAULT, ///< Default priority used, e.g. the default timer priority
invoke idle priority timers only when actually idle The 'Idle' timers are misnamed. They are zero-timeout times, i.e. they are invoked immediately after returning to the main loop. But that does not necessarily mean they are invoked when idle, there may be e.g. user input pending in the system event queue. In fact, LO events are processed before system events, which means that 'Idle' timers are normally processed before user input. Besides being confused, this also leads to poor performance in some cases, such as when using mouse wheel to zoom in a large document. This results in several mouse wheel events, each of which will result in adjusting the zoom and that causing a repaint. Repaints are internally handled using a TaskPriority::REPAINT 'Idle', and so what happens is zoom->repaint->zoom->repaint->zoom->repaint instead of the more efficient zoom->zoom->zoom->repaint. This change (besides trying to clarify the confusion in the docs) delays invoking tasks with priorities TaskPriority::HIGH_IDLE and lower if there is user input or repaint events in the OS queue. That means that tasks using idle priorities actually will be invoked only when idle (barring background threads etc.). I'm reasonably certain this is a safe change, there's no guarantee when exactly tasks will be invoked (e.g. other tasks with a higher priority go first) and explicitly specifying such a priority means asking for it. I already implemented this once in 06d731428ef6cf93c7333e8228b, and it was also again done in 87199d3829257420429057336283, but apparently these have been removed. There was d348035a60361a1b9ba9e 'Drop special idle handling' with the reasoning that 'Idles are just instant timers'. Which strictly technically speaking is true due to 'Idle' being a misnomer, but the point is that some idles should be actual idles and that's why they need to be handled specially. Change-Id: I36c2b02a80ae7e1476b731f878d9b28aa87975f4 Reviewed-on: https://gerrit.libreoffice.org/c/core/+/110538 Tested-by: Jenkins Reviewed-by: Luboš Luňák <l.lunak@collabora.com>
2021-02-07 16:34:38 +01:00
// Input from the OS event queue is processed before HIGH_IDLE tasks.
HIGH_IDLE, ///< Important idle events to be run before processing drawing events
RESIZE, ///< Resize runs before repaint, so we won't paint twice
REPAINT, ///< All repaint events should go in here
tdf#165277 On macOS, only delay priorities lower than POST_PAINT macOS bugs tdf#157312 and tdf#163945 were fixed by firing the Skia flush task with TaskPriority::POST_PAINT. The problem is that this method often executes within an NSTimer and NSTimers are always fired while LibreOffice is in -[NSApp nextEventMatchingMask:untilDate:inMode:dequeue:]. Since fetching the next native event doesn't handle pending events until *after* all of the pending NSTimers have fired, calling SalInstance::AnyInput() will almost always return true due to the pending events that will be handled immediately after all of the the pending NSTimers have fired. The result is that the Skia flush task is frequently delayed and, in cases like tdf#165277, a user's attempts to get LibreOffice to paint the window through key and mouse events leads to an endless delaying of the Skia flush task. After experimenting with both Skia/Metal and Skia/Raster, tdf#165277 requires the Skia flush task to run immediately before the TaskPriority::POST_PAINT tasks. After that, all TaskPriority::POST_PAINT tasks must run so the Skia flush task now uses the TaskPriority::SKIA_FLUSH priority on macOS. One positive side effect of this change is that live resizing on macOS is now much smoother. Even with Skia disabled (which does not paint using a task but does use tasks to handle live resizing), the content resizes much more quickly when a user repidly changes window's size. Change-Id: I727b2d215382dbd2ebc2608d401f8b5c0dd49773 Reviewed-on: https://gerrit.libreoffice.org/c/core/+/182380 Reviewed-by: Patrick Luby <guibomacdev@gmail.com> Tested-by: Jenkins
2025-02-28 22:52:13 -05:00
SKIA_FLUSH, ///< tdf#165277 Skia needs to flush immediately before POST_PAINT tasks on macOS
POST_PAINT, ///< Everything running directly after painting
DEFAULT_IDLE, ///< Default idle priority
LOWEST, ///< Low, very idle cleanup tasks
TOOLKIT_DEBUG ///< Do not use. Solely for IdleTask::waitUntilIdleDispatched
};
#define PRIO_COUNT (static_cast<int>(TaskPriority::TOOLKIT_DEBUG) + 1)
class VCL_DLLPUBLIC Task
{
friend class Scheduler;
friend struct ImplSchedulerData;
ImplSchedulerData *mpSchedulerData; ///< Pointer to the element in scheduler list
const char *mpDebugName; ///< Useful for debugging
TaskPriority mePriority; ///< Task priority
bool mbActive; ///< Currently in the scheduler
bool mbStatic; ///< Is a static object
protected:
static void StartTimer( sal_uInt64 nMS );
const ImplSchedulerData* GetSchedulerData() const { return mpSchedulerData; }
virtual void SetDeletionFlags();
/**
* How long (in MS) until the Task is ready to be dispatched?
*
* Simply return Scheduler::ImmediateTimeoutMs if you're ready, like an
* Idle. If you have to return Scheduler::InfiniteTimeoutMs, you probably
* need another mechanism to wake up the Scheduler or rely on other
* Tasks to be scheduled, or simply use a polling Timer.
*
* @param nTimeNow the current time
* @return the sleep time of the Task to become ready
*/
virtual sal_uInt64 UpdateMinPeriod( sal_uInt64 nTimeNow ) const = 0;
public:
Task( const char *pDebugName );
Task( const Task& rTask );
virtual ~Task() COVERITY_NOEXCEPT_FALSE;
Task& operator=( const Task& rTask );
void SetPriority(TaskPriority ePriority);
TaskPriority GetPriority() const { return mePriority; }
const char *GetDebugName() const { return mpDebugName; }
Experimental Emscripten Qt6 --enable-emscripten-proxy-to-pthread ersatz With Qt5, the (enabled by default) --enable-emscripten-proxy-to-pthread (i.e., leveraging Emscripten's -sPROXY_TO_PTHREAD) works reasonably well with the handful of hacks we added to our Qt5 5.15.2+wasm branch. However, for Qt6 any necessary hacking there would need to be excessive and looks unrealistic (see the mailing list thread starting at <https://lists.qt-project.org/pipermail/development/2024-December/045960.html> "[Development] Wasm: Support for Emscripten PROXY_TO_PTHREAD?" and the presentation at <https://fosdem.org/2025/schedule/event/fosdem-2025-5169-lowa-in-need-of-a-vcl-plug/>). But when leveraging the upcoming browser support for JSPI (through Emscripten's -sJSPI, Qt6's -feature-wasm-jspi, and our --enable-emscripten-jspi), it looks feasible to leave the Qt6 code alone and instead of proxying the complete application main thread (including the Qt6 event handling) off the browser main thread to an additional thread (automatically set up through Emscripten -sPROXY_TO_PTHREAD), to just proxy some of the LO event handling off the browser main thread to an additional thread (set up through a new comphelper::emscriptenthreading::setUp). That way, problematic LO code that is triggered from event handling (like "Tools - Extensions..." wanting to spawn additional threads, which requires the browser main thread to be unblocked) can be offloaded to the additional thread and get the browser main thread unblocked. Thus there is a new "Emscripten Qt6 JSPI/non-PROXY_TO_PTHREAD" mode now (via --enable-qt6 --enable-emscripten-jspi --disable-emscripten-proxy-to-pthread). It is still experimental and known to occasionally hang and crash, but has been seen to generally work at least with recent emsdk 4.0.3 and a recent Qt dev branch (towards Qt 6.10, built with -feature-wasm-jspi) and running on recent Chrome 132 with "Experimental WebAssembly JavaScript Promise Integration (JSPI)" enabled under <chrome://flags>. There are two places where code is proxied off the browser main thread to the additional thread: One is QtInstance::ProcessEvent, the other is invoking certain kinds of tasks (just "vcl::Dialog maLayoutIdle", for now) in Scheduler::CallbackTaskScheduling (unconditionally proxying off all kinds of tasks there appeared to cause more issues than it would solve). For the latter, the concept of "transferability" has been added to class Task; the new Emscripten mode is the only case making use of that concept. Also for the latter, it is important that the browser main thread does not get blocked at the > SolarMutexGuard aGuard; in QtTimer::timeoutActivated when such a timer event happens while the additional thread has the SolarMutex locked and tries to do something that requires the browser main thread to not be blocked (like spawning additional threads upon "Tools - Extensions..."). Therefore, as a bad hack, and only for the new Emscripen mode, that SolarMutexGuard has been moved from QtTimer::timeoutActivated down to the call of pTask->Invoke() in Scheduler::CallbackTaskScheduling (see the two TODO comments added to the code). This appears to work relatively well, but is still a bad hack that should rather be done properly. And then there's some code across vcl/qt5 that may now end up running proxied off the browser main thread, but which actually requires to be run on the browser main thread (because, e.g., it calls into Qt6 code that accesses JS entities that are only available there). Those places use a newly introduced QtInstance::EmscriptenLightweightRunInMainThread that takes some code via a lambda: For the new Emscripten mode, the code is explicitly proxied back to the browser main thread, while in all other cases it is run directly in place. Some further notes: * -sPTHREAD_POOL_SIZE needed to be bumped by one, to accommodate for the additional new thread. * The code requires two additional JSPI "entry points", _emscripten_check_mailbox (to complement the suspension in QtInstance::ProcessEvent) and the invocation of the Qt6-internal qstdweb::EventListener::handleEvent (to complement a Qt6-internal suspension point hit when e.g. doing "Tools - Extension Manager..." and clicking "Add"). * The new Emscripten mode requires -fexperimental-library (for std::jthread and std::stop_token) when building at least with emsdk 4.0.3, but it shouldn't hurt to have that switch enabled unconditionally for Emscripten builds (which are self-contained, so can't be hit by any compatibility issues that switch might cause). Change-Id: Id3b13e2cc3c304b072f388c6d736e7e663b36c58 Reviewed-on: https://gerrit.libreoffice.org/c/core/+/181424 Tested-by: Jenkins Reviewed-by: Stephan Bergmann <stephan.bergmann@allotropia.de> Reviewed-by: Michael Weghorn <m.weghorn@posteo.de>
2025-02-11 14:11:25 +01:00
virtual bool DecideTransferredExecution();
// Call handler
virtual void Invoke() = 0;
/**
* Schedules the task for execution
*
* If the timer is already active, it's reset!
* Check with Task::IsActive() to prevent reset.
*
* If you unset bStartTimer, the Task must call Task::StartTimer(...) to be correctly scheduled!
* Otherwise it might just be picked up when the Scheduler runs the next time.
*
* @param bStartTimer if false, don't schedule the Task by calling Task::StartTimer(0).
*/
virtual void Start(bool bStartTimer = true);
void Stop();
bool IsActive() const { return mbActive; }
/**
* This function must be called for static tasks, so the Task destructor
* ignores the scheduler mutex, as it may not be available anymore.
* The cleanup is still correct, as it has already happened in
* DeInitScheduler call well before the static destructor calls.
*/
void SetStatic() { mbStatic = true; }
bool IsStatic() const { return mbStatic; }
};
#endif // INCLUDED_VCL_TASK_HXX
/* vim:set shiftwidth=4 softtabstop=4 expandtab: */