ferroui_base/rendering/composition/

compositor.rs

use super::server::{BatchQueue, CompositorClock, ServerCompositionVisual, ServerCompositor, ServerObjectId};
use super::transport::{
    BatchMarker, BatchObject, BatchStreamWriter, CommittedBatch, CompositionBatch, ServerJob, ServerObjectFactory, ServerObjectJob,
};
use super::{CompositionOptions, CompositionVisual, ICompositorSerializable};
use crate::animation::easings::{IEasing, SplineEasing};
use crate::animation::KeySpline;
use crate::media::imaging::Bitmap;
use crate::platform::IPlatformGraphics;
use crate::reactive::{Disposable, IDisposable};
use crate::rendering::composition::server::IServerObject;
use crate::rendering::{IRenderLoop, IRenderLoopTask};
use crate::threading::{Dispatcher, DispatcherPriority};
use crate::utilities::HandlerList;
use crate::{FerroLocator, LocatorExtensions};
use std::cell::{Cell, RefCell};
use std::collections::{HashMap, HashSet, VecDeque};
use std::rc::{Rc, Weak};
use std::sync::atomic::{AtomicBool, AtomicU64, Ordering};
use std::sync::{Arc, Mutex};
use std::time::Duration;

/// Receives the request to commit a compositor's pending changes; the media
/// context implements it and schedules the commit with the next frame.
pub trait ICompositorScheduler {
    fn commit_requested(&self, compositor: &Rc<Compositor>);
}

static NEXT_COMPOSITOR_KEY: AtomicU64 = AtomicU64::new(1);

crate::ferro_thread_local! {
    /// The compositors of this thread by key: lets `Send` callbacks that
    /// come back to the thread find their compositor.
    static COMPOSITORS: RefCell<HashMap<u64, Weak<Compositor>>> = RefCell::new(HashMap::new());
}

fn find_compositor(key: u64) -> Option<Rc<Compositor>> {
    COMPOSITORS.with(|compositors| compositors.borrow().get(&key).and_then(Weak::upgrade))
}

/// The compositor: manages the composition objects of the UI thread and
/// commits their changes to the server compositor in batches.
pub struct Compositor {
    this: Weak<Compositor>,
    key: u64,
    render_loop: Arc<dyn IRenderLoop>,
    loop_task: Arc<dyn IRenderLoopTask>,
    use_ui_thread_for_synchronous_commits: bool,
    /// The server compositor, behind the compositor lock.
    server: Arc<super::server::LockedServerCompositor>,
    /// Whether the render loop renders on the thread that ticks it (the
    /// render-thread mode); otherwise a tick is marshalled to the thread of
    /// this compositor (the dispatcher-thread mode).
    render_thread: bool,
    /// The loop task of the render-thread mode when the server compositor is
    /// confined to the render thread (see
    /// [`is_confined_to_render_thread`](Self::is_confined_to_render_thread)):
    /// the compositor asks it for the release of the server graph.
    confined_loop_task: Option<Arc<RenderThreadLoopTask>>,
    /// Whether a job of the render thread is on its way to fill the cache
    /// of the features of the render interface.
    /// The readback of the server: what the two threads share besides the
    /// queue of batches.
    readback: Arc<super::server::ReadbackIndices>,
    batches: Arc<BatchQueue>,
    clock: CompositorClock,
    next_commit: RefCell<Option<Arc<CompositionBatch>>>,
    object_serialization_queue: RefCell<VecDeque<Rc<dyn ICompositorSerializable>>>,
    object_serialization_hash_set: RefCell<HashSet<*const ()>>,
    invoke_before_commit_write: RefCell<VecDeque<Box<dyn FnOnce()>>>,
    invoke_before_commit_read: RefCell<VecDeque<Box<dyn FnOnce()>>>,
    dispose_on_next_batch: RefCell<Vec<PendingDisposal>>,
    pending_creations: RefCell<Vec<(ServerObjectId, ServerObjectFactory)>>,
    next_server_object_id: Cell<u32>,
    /// The ids that can be handed out again: each was released by a batch
    /// that is committed, so the server drops its object before it reads a
    /// creation under the same id, which can only come with a later batch.
    free_server_object_ids: RefCell<Vec<u32>>,
    /// By id, whether the id is handed out: an id is returned to the free
    /// list once, however often its release is asked for.
    server_object_id_in_use: RefCell<Vec<bool>>,
    pending_batch: Arc<Mutex<Option<Arc<CompositionBatch>>>>,
    pending_server_compositor_jobs: RefCell<Vec<PendingServerJob>>,
    pending_server_compositor_post_target_jobs: RefCell<Vec<PendingServerJob>>,
    scheduler: Weak<dyn ICompositorScheduler>,
    dispatcher: Arc<Dispatcher>,
    after_commit: HandlerList<dyn Fn()>,
    default_easing: Rc<dyn IEasing>,
    /// The open scoped batches and the batch of the current frame (WinUI
    /// `CompositionScopedBatch` and `CompositionCommitBatch`).
    animation_batches: super::composition_batch::CompositorBatches,
}

/// The render loop task of a compositor. The loop may tick on any thread;
/// the server compositor is confined to the thread of its compositor, so a
/// tick from elsewhere is marshalled there.
struct ServerCompositorLoopTask {
    key: u64,
    dispatcher: Arc<Dispatcher>,
    marshalled: Arc<AtomicBool>,
    last_result: Arc<AtomicBool>,
}

impl IRenderLoopTask for ServerCompositorLoopTask {
    fn render(&self) -> bool {
        if let Some(compositor) = find_compositor(self.key) {
            let result = compositor.server().render();
            self.last_result.store(result, Ordering::SeqCst);
            return result;
        }
        if !self.marshalled.swap(true, Ordering::SeqCst) {
            let key = self.key;
            let marshalled = self.marshalled.clone();
            let last_result = self.last_result.clone();
            self.dispatcher.post(
                move || {
                    marshalled.store(false, Ordering::SeqCst);
                    if let Some(compositor) = find_compositor(key) {
                        last_result.store(compositor.server().render(), Ordering::SeqCst);
                    }
                },
                DispatcherPriority::RENDER,
            );
        }
        self.last_result.load(Ordering::SeqCst)
    }
}

/// The clock of a compositor whose server runs on the render thread: read
/// by both threads.
pub type SharedCompositorClock = Arc<dyn Fn() -> Duration + Send + Sync>;

/// The render loop task of a compositor in the render-thread mode: the
/// thread that ticks renders, under the compositor lock.
struct RenderThreadLoopTask {
    this: std::sync::Weak<RenderThreadLoopTask>,
    server: Arc<super::server::LockedServerCompositor>,
    /// Whether the thread that ticks is the only one that renders: the
    /// server graph is then confined to it, and released by it.
    confined: bool,
    /// Set by the compositor when it is dropped, where `confined`: the next
    /// tick releases the server graph instead of rendering.
    release_requested: AtomicBool,
    removal_posted: AtomicBool,
    render_loop: std::sync::Weak<dyn IRenderLoop>,
    dispatcher: Arc<Dispatcher>,
}

impl RenderThreadLoopTask {
    /// Asks the render thread to release the server graph with its next
    /// tick. From then on no frame is rendered.
    fn request_release(&self) {
        self.release_requested.store(true, Ordering::SeqCst);
    }

    /// Takes the task out of the render loop once the graph is released. A
    /// task is added and removed by the thread of the compositor, so the
    /// removal is posted there.
    fn remove_from_loop(&self) {
        if self.removal_posted.swap(true, Ordering::SeqCst) {
            return;
        }
        let Some(this) = self.this.upgrade() else { return };
        let render_loop = self.render_loop.clone();
        self.dispatcher.post(
            move || {
                if let Some(render_loop) = render_loop.upgrade() {
                    let task: Arc<dyn IRenderLoopTask> = this;
                    render_loop.remove(&task);
                }
            },
            DispatcherPriority::SEND,
        );
    }
}

impl IRenderLoopTask for RenderThreadLoopTask {
    fn render(&self) -> bool {
        if self.confined {
            self.server.confine_to_current_thread();
        }
        // The request is read inside the lock: the compositor sets it before
        // it enters the lock to take its handles back, so a frame either
        // ends before that or does not start.
        let rendered = self
            .server
            .try_with(|server| (!self.release_requested.load(Ordering::SeqCst)).then(|| server.render()));
        match rendered {
            Some(Some(wants_next_tick)) => wants_next_tick,
            Some(None) => {
                // The last job of the render thread: the graph, and with it
                // every render target and the graphics context, is dropped
                // here. What it drew with is released first, while its own
                // graphics context is current: a thread that renders several
                // compositors has several contexts, and between two frames
                // any of them, or none, is the current one.
                self.server.try_with(|server| server.release_gpu_resources());
                self.server.release();
                self.remove_from_loop();
                false
            }
            // Nothing to render once the compositor is gone.
            None => false,
        }
    }
}

impl Compositor {
    /// Creates a compositor in the render-thread mode: the thread that ticks
    /// `render_loop` renders the frames.
    ///
    /// The server compositor is confined to the compositor lock, not to a
    /// thread: with `use_ui_thread_for_synchronous_commits` the thread of
    /// this compositor renders too, at the synchronous points (a resize, the
    /// first show), as the reference does on platforms that ask for it.
    ///
    /// Without it, and with a loop that runs in the background, the server
    /// compositor is also confined to the thread that ticks (see
    /// [`is_confined_to_render_thread`](Self::is_confined_to_render_thread)).
    /// The loop says where it runs when the compositor is created.
    ///
    /// `gpu` is shared with the caller, who may keep its handle: the server
    /// side holds one of its own and the thread that renders asks it for
    /// the graphics context.
    pub fn with_render_thread(
        render_loop: Arc<dyn IRenderLoop>,
        gpu: Option<Arc<dyn IPlatformGraphics>>,
        use_ui_thread_for_synchronous_commits: bool,
        scheduler: &Rc<dyn ICompositorScheduler>,
        dispatcher: Arc<Dispatcher>,
        options: Option<CompositionOptions>,
        clock: Option<SharedCompositorClock>,
    ) -> Rc<Compositor> {
        let options = options
            .or_else(|| FerroLocator::current().get_service::<CompositionOptions>().map(|o| *o))
            .unwrap_or_default();
        // The clock is read by both threads. Each side has a handle of its
        // own to it: the server graph may be released by the render thread.
        let clock = clock.unwrap_or_else(|| {
            let origin = std::time::Instant::now();
            Arc::new(move || origin.elapsed())
        });
        let server_clock: CompositorClock = {
            let clock = clock.clone();
            Rc::new(move || clock())
        };
        let clock: CompositorClock = Rc::new(move || clock());
        let batches = Arc::new(BatchQueue::default());
        let readback = Arc::new(super::server::ReadbackIndices::new());
        let server = Arc::new(super::server::LockedServerCompositor::new(ServerCompositor::new(
            gpu,
            options,
            batches.clone(),
            readback.clone(),
            server_clock,
        )));
        let key = NEXT_COMPOSITOR_KEY.fetch_add(1, Ordering::SeqCst);
        // The condition of the reference for a thread of the compositor that
        // never renders (`UseUiThreadForSynchronousCommits: false`,
        // `Loop.RunsInBackground: true`).
        let confined = !use_ui_thread_for_synchronous_commits && render_loop.runs_in_background();
        let task = Arc::new_cyclic(|this| RenderThreadLoopTask {
            this: this.clone(),
            server: server.clone(),
            confined,
            release_requested: AtomicBool::new(false),
            removal_posted: AtomicBool::new(false),
            render_loop: Arc::downgrade(&render_loop),
            dispatcher: dispatcher.clone(),
        });
        let loop_task: Arc<dyn IRenderLoopTask> = task.clone();
        Self::build(
            render_loop,
            loop_task,
            confined.then_some(task),
            use_ui_thread_for_synchronous_commits,
            server,
            true,
            batches,
            readback,
            clock,
            scheduler,
            dispatcher,
            key,
        )
    }

    /// Creates a compositor that renders on the render loop registered in
    /// the service locator and commits through the media context.
    ///
    /// The threads follow the render loop, as in the reference: with a loop
    /// that runs in the background (the thread of the platform's render
    /// timer) that thread renders the frames, under the compositor lock (the
    /// render-thread mode); with a loop of the UI thread the UI thread
    /// renders. `FERROUI_RENDER_THREAD=0` in the environment keeps the
    /// rendering on the UI thread with a background loop too (each tick is
    /// marshalled to it): a way back while the mode is young.
    pub fn new(gpu: Option<Arc<dyn IPlatformGraphics>>, use_ui_thread_for_synchronous_commits: bool) -> Rc<Compositor> {
        let render_loop = FerroLocator::current().get_required_service::<Arc<dyn IRenderLoop>>();
        let scheduler: Rc<dyn ICompositorScheduler> = crate::media::MediaContext::instance().scheduler();
        let render_thread_disabled = std::env::var("FERROUI_RENDER_THREAD").is_ok_and(|value| value == "0");
        if render_loop.runs_in_background() && !render_thread_disabled {
            return Self::with_render_thread(
                (*render_loop).clone(),
                gpu,
                use_ui_thread_for_synchronous_commits,
                &scheduler,
                Dispatcher::ui_thread(),
                None,
                None,
            );
        }
        Self::with_scheduler(
            (*render_loop).clone(),
            gpu,
            use_ui_thread_for_synchronous_commits,
            &scheduler,
            Dispatcher::ui_thread(),
            None,
            None,
        )
    }

    /// Creates a compositor with every collaborator given explicitly.
    ///
    /// `options` defaults to the options registered in the service locator;
    /// `clock` defaults to the dispatcher's clock.
    pub fn with_scheduler(
        render_loop: Arc<dyn IRenderLoop>,
        gpu: Option<Arc<dyn IPlatformGraphics>>,
        use_ui_thread_for_synchronous_commits: bool,
        scheduler: &Rc<dyn ICompositorScheduler>,
        dispatcher: Arc<Dispatcher>,
        options: Option<CompositionOptions>,
        clock: Option<CompositorClock>,
    ) -> Rc<Compositor> {
        let options = options
            .or_else(|| FerroLocator::current().get_service::<CompositionOptions>().map(|o| *o))
            .unwrap_or_default();
        let clock = clock.unwrap_or_else(|| {
            let platform_impl = dispatcher.platform_impl();
            let origin = platform_impl.now();
            Rc::new(move || Duration::from_millis((platform_impl.now() - origin).max(0) as u64))
        });
        let batches = Arc::new(BatchQueue::default());
        let readback = Arc::new(super::server::ReadbackIndices::new());
        let server = Arc::new(super::server::LockedServerCompositor::new(ServerCompositor::new(
            gpu,
            options,
            batches.clone(),
            readback.clone(),
            clock.clone(),
        )));
        let key = NEXT_COMPOSITOR_KEY.fetch_add(1, Ordering::SeqCst);
        let loop_task: Arc<dyn IRenderLoopTask> = Arc::new(ServerCompositorLoopTask {
            key,
            dispatcher: dispatcher.clone(),
            marshalled: Arc::new(AtomicBool::new(false)),
            last_result: Arc::new(AtomicBool::new(true)),
        });
        Self::build(
            render_loop,
            loop_task,
            None,
            use_ui_thread_for_synchronous_commits,
            server,
            false,
            batches,
            readback,
            clock,
            scheduler,
            dispatcher,
            key,
        )
    }

    #[allow(clippy::too_many_arguments)]
    fn build(
        render_loop: Arc<dyn IRenderLoop>,
        loop_task: Arc<dyn IRenderLoopTask>,
        confined_loop_task: Option<Arc<RenderThreadLoopTask>>,
        use_ui_thread_for_synchronous_commits: bool,
        server: Arc<super::server::LockedServerCompositor>,
        render_thread: bool,
        batches: Arc<BatchQueue>,
        readback: Arc<super::server::ReadbackIndices>,
        clock: CompositorClock,
        scheduler: &Rc<dyn ICompositorScheduler>,
        dispatcher: Arc<Dispatcher>,
        key: u64,
    ) -> Rc<Compositor> {
        let compositor = Rc::new_cyclic(|this| Compositor {
            this: this.clone(),
            key,
            render_loop: render_loop.clone(),
            loop_task: loop_task.clone(),
            use_ui_thread_for_synchronous_commits,
            readback,
            server,
            render_thread,
            confined_loop_task,
            batches,
            clock,
            next_commit: RefCell::new(None),
            object_serialization_queue: RefCell::new(VecDeque::new()),
            object_serialization_hash_set: RefCell::new(HashSet::new()),
            invoke_before_commit_write: RefCell::new(VecDeque::new()),
            invoke_before_commit_read: RefCell::new(VecDeque::new()),
            dispose_on_next_batch: RefCell::new(Vec::new()),
            pending_creations: RefCell::new(Vec::new()),
            next_server_object_id: Cell::new(0),
            free_server_object_ids: RefCell::new(Vec::new()),
            server_object_id_in_use: RefCell::new(Vec::new()),
            pending_batch: Arc::new(Mutex::new(None)),
            pending_server_compositor_jobs: RefCell::new(Vec::new()),
            pending_server_compositor_post_target_jobs: RefCell::new(Vec::new()),
            scheduler: Rc::downgrade(scheduler),
            dispatcher,
            after_commit: HandlerList::new(),
            default_easing: Rc::new(SplineEasing::with_key_spline(KeySpline::with_points(0.25, 0.1, 0.25, 1.0))),
            animation_batches: super::composition_batch::CompositorBatches::new(),
        });
        COMPOSITORS.with(|compositors| {
            let mut compositors = compositors.borrow_mut();
            compositors.retain(|_, c| c.strong_count() > 0);
            compositors.insert(key, Rc::downgrade(&compositor));
        });
        render_loop.add(loop_task);
        compositor
    }

    /// The easing of the key frames that are inserted without one.
    pub(crate) fn default_easing(&self) -> Rc<dyn IEasing> {
        self.default_easing.clone()
    }

    /// The open scoped batches and the batch of the current frame.
    pub(crate) fn animation_batches(&self) -> &super::composition_batch::CompositorBatches {
        &self.animation_batches
    }

    /// The render loop the server compositor is driven by.
    pub fn render_loop(&self) -> &Arc<dyn IRenderLoop> {
        &self.render_loop
    }

    pub fn use_ui_thread_for_synchronous_commits(&self) -> bool {
        self.use_ui_thread_for_synchronous_commits
    }

    /// The server compositor. It belongs to the render thread: UI-thread
    /// code may only use it to drive a frame when the render loop runs on
    /// the UI thread, and in tests.
    pub fn server(&self) -> &Rc<ServerCompositor> {
        assert!(
            !self.render_thread,
            "in the render-thread mode the server compositor is reached under its lock: with_server"
        );
        self.server.same_thread()
    }

    /// Runs `f` with the server compositor under the compositor lock: the
    /// way in from this thread in either mode. What `f` gets must not leave
    /// it.
    pub fn with_server<R>(&self, f: impl FnOnce(&Rc<ServerCompositor>) -> R) -> R {
        self.server.with(f)
    }

    /// The server compositor behind its lock, as a handle that may cross to
    /// the thread that renders: for work of that thread that cannot be a job
    /// of a batch, because it has to run whether or not this compositor
    /// commits again and whether or not its graphics are ready (the release
    /// of a surface that is closed). The work enters the graph with
    /// [`try_with`](super::server::LockedServerCompositor::try_with), which
    /// answers `None` once the graph is released.
    ///
    /// Not from upstream, where such work captures the server compositor
    /// itself.
    pub fn locked_server(&self) -> Arc<super::server::LockedServerCompositor> {
        self.server.clone()
    }

    /// Binds `value` to the compositor lock: an object of the server side
    /// that this thread asks for and jobs of the render thread use (see
    /// [`LockBound`](super::server::LockBound)). `server` is what
    /// [`with_server`](Self::with_server) or a job hands out.
    pub(crate) fn bind_to_lock<T>(&self, server: &ServerCompositor, value: T) -> super::server::LockBound<T> {
        super::server::LockBound::new(&self.server, server, value)
    }

    /// Renders a frame on this thread, under the compositor lock: what the
    /// synchronous points do (`Server.Render` upstream).
    ///
    /// # Panics
    ///
    /// When the compositor is confined to its render thread: this thread
    /// waits for a frame there, it never renders one.
    pub fn render_on_this_thread(&self) -> bool {
        assert!(
            !self.is_confined_to_render_thread(),
            "a compositor that is confined to its render thread is not rendered by the thread of the compositor"
        );
        self.server.with(|server| server.render())
    }

    /// Whether the thread that ticks the render loop renders the frames,
    /// not the thread of this compositor.
    pub fn renders_on_render_thread(&self) -> bool {
        self.render_thread
    }

    /// Whether the thread that ticks the render loop is the only one that
    /// renders: the render-thread mode without
    /// `use_ui_thread_for_synchronous_commits`, on a loop that runs in the
    /// background (the condition under which the reference waits for the
    /// render thread at the synchronous points).
    ///
    /// The render targets, the graphics context and everything else that
    /// belongs to the thread that draws are then created, used and dropped
    /// by the render thread alone. This thread still enters the compositor
    /// lock ([`with_server`](Self::with_server)), for what does not render:
    /// it commits batches, posts jobs and waits for them.
    pub fn is_confined_to_render_thread(&self) -> bool {
        self.confined_loop_task.is_some()
    }

    /// The readback indices the server writes and this side reads.
    pub fn readback(&self) -> &Arc<super::server::ReadbackIndices> {
        &self.readback
    }

    /// The dispatcher of the thread the compositor belongs to.
    pub fn dispatcher(&self) -> &Arc<Dispatcher> {
        &self.dispatcher
    }

    /// The time elapsed on the compositor clock.
    pub fn clock_elapsed(&self) -> Duration {
        (self.clock)()
    }

    /// Subscribes to the notification raised after each commit.
    pub fn after_commit(&self, handler: impl Fn() + 'static) -> Rc<dyn IDisposable> {
        let token = self.after_commit.add(Rc::new(handler));
        let weak = self.this.clone();
        Disposable::create(move || {
            if let Some(this) = weak.upgrade() {
                this.after_commit.remove(token);
            }
        })
    }

    pub(crate) fn this_handle(&self) -> Rc<Compositor> {
        self.this()
    }

    fn this(&self) -> Rc<Compositor> {
        self.this.upgrade().expect("the compositor is alive while it is used")
    }

    fn trigger_commit_requested(&self) {
        if let Some(scheduler) = self.scheduler.upgrade() {
            scheduler.commit_requested(&self.this());
        }
    }

    /// Requests the pending changes to be committed and returns the batch
    /// they will be part of.
    pub fn request_composition_batch_commit_async(&self) -> Arc<CompositionBatch> {
        self.dispatcher.verify_access();
        if let Some(next_commit) = self.next_commit.borrow().clone() {
            return next_commit;
        }
        let next_commit = CompositionBatch::new();
        *self.next_commit.borrow_mut() = Some(next_commit.clone());
        let pending = self.pending_batch.lock().unwrap_or_else(|e| e.into_inner()).clone();
        match pending {
            Some(pending) => {
                // The previous batch has not been processed yet: ask for
                // the commit once it has been.
                let key = self.key;
                let dispatcher = self.dispatcher.clone();
                pending.processed().on_completed(move || {
                    dispatcher.post(
                        move || {
                            if let Some(compositor) = find_compositor(key) {
                                compositor.trigger_commit_requested();
                            }
                        },
                        DispatcherPriority::SEND,
                    );
                });
            }
            None => self.trigger_commit_requested(),
        }
        next_commit
    }

    /// Requests the pending changes to be committed.
    pub fn request_commit_async(&self) -> Arc<CompositionBatch> {
        self.request_composition_batch_commit_async()
    }

    /// Serializes the pending changes into a batch and hands it to the
    /// server compositor.
    pub fn commit(&self) -> Arc<CompositionBatch> {
        let result = self.commit_core();
        if !self.invoke_before_commit_write.borrow().is_empty() {
            self.request_commit_async();
        }
        if !self.after_commit.is_empty() {
            for (_, handler) in self.after_commit.snapshot().iter() {
                handler();
            }
        }
        result
    }

    fn commit_core(&self) -> Arc<CompositionBatch> {
        self.dispatcher.verify_access();

        let commit = self.next_commit.borrow_mut().get_or_insert_with(CompositionBatch::new).clone();

        std::mem::swap(
            &mut *self.invoke_before_commit_read.borrow_mut(),
            &mut *self.invoke_before_commit_write.borrow_mut(),
        );
        loop {
            let Some(action) = self.invoke_before_commit_read.borrow_mut().pop_front() else { break };
            action();
        }

        let mut changes = self.batches.rent_data();
        let disposed: Vec<PendingDisposal>;
        {
            let mut writer = BatchStreamWriter::new(&mut changes);

            // Objects serialize references to server objects, so the objects
            // created since the last commit come first.
            let mut objects_written = false;
            // The set of queued objects knows an object by its address, so
            // an object stays alive until the set is cleared: an object
            // created while another one is serialized can then never get the
            // address of one that was serialized (and released) before it,
            // and be taken for queued.
            let mut serialized: Vec<Rc<dyn ICompositorSerializable>> = Vec::new();
            loop {
                self.write_pending_creations(&mut writer);
                let Some(object) = self.object_serialization_queue.borrow_mut().pop_front() else { break };
                if let Some(server_object) = object.try_get_server(self) {
                    // Server objects the object creates for its changes
                    // must exist before the changes are read.
                    object.prepare_serialization(self);
                    self.write_pending_creations(&mut writer);
                    writer.write_server_object(Some(server_object));
                    object.serialize_changes(self, &mut writer);
                    if cfg!(debug_assertions) {
                        writer.write(super::server::OBJECT_END_MAGIC);
                        writer.write_marker(BatchMarker::ObjectEnd);
                    }
                    objects_written = true;
                }
                serialized.push(object);
            }
            let _ = objects_written;
            self.object_serialization_hash_set.borrow_mut().clear();
            drop(serialized);

            disposed = std::mem::take(&mut *self.dispose_on_next_batch.borrow_mut());
            if !disposed.is_empty() {
                writer.write_marker(BatchMarker::RenderThreadDisposeStart);
                writer.write(disposed.len() as i32);
                for entry in &disposed {
                    writer.write_server_object(Some(entry.id));
                    writer.write(entry.dispose);
                    writer.write(entry.release);
                }
            }

            Self::serialize_server_jobs(
                &mut writer,
                &self.pending_server_compositor_jobs,
                BatchMarker::RenderThreadJobsStart,
                BatchMarker::RenderThreadJobsEnd,
            );
            Self::serialize_server_jobs(
                &mut writer,
                &self.pending_server_compositor_post_target_jobs,
                BatchMarker::RenderThreadPostTargetJobsStart,
                BatchMarker::RenderThreadPostTargetJobsEnd,
            );
        }
        // Ids released by this batch can be reused by objects created for a
        // later one.
        {
            let mut in_use = self.server_object_id_in_use.borrow_mut();
            let mut free = self.free_server_object_ids.borrow_mut();
            for entry in disposed.iter().filter(|entry| entry.release) {
                if in_use.get_mut(entry.id.index()).is_some_and(|in_use| std::mem::replace(in_use, false)) {
                    free.push(entry.id.0);
                }
            }
        }

        self.enqueue_batch(CommittedBatch { batch: commit.clone(), changes, committed_at: self.clock_elapsed() });

        *self.pending_batch.lock().unwrap_or_else(|e| e.into_inner()) = Some(commit.clone());
        let pending_batch = self.pending_batch.clone();
        let sequence_id = commit.sequence_id();
        commit.processed().on_completed(move || {
            let mut pending = pending_batch.lock().unwrap_or_else(|e| e.into_inner());
            if pending.as_ref().is_some_and(|p| p.sequence_id() == sequence_id) {
                *pending = None;
            }
        });
        *self.next_commit.borrow_mut() = None;
        // The animations started in the frame are on their way: the commit
        // batch of the frame is ended.
        self.animation_batches.on_commit();

        commit
    }

    fn write_pending_creations(&self, writer: &mut BatchStreamWriter<'_>) {
        let creations = std::mem::take(&mut *self.pending_creations.borrow_mut());
        if creations.is_empty() {
            return;
        }
        writer.write_marker(BatchMarker::CreateStart);
        writer.write(creations.len() as i32);
        for (id, factory) in creations {
            writer.write_server_object(Some(id));
            writer.write_object(BatchObject::Create(factory));
        }
    }

    fn serialize_server_jobs(
        writer: &mut BatchStreamWriter<'_>,
        list: &RefCell<Vec<PendingServerJob>>,
        start_marker: BatchMarker,
        end_marker: BatchMarker,
    ) {
        let jobs = std::mem::take(&mut *list.borrow_mut());
        if !jobs.is_empty() {
            writer.write_marker(start_marker);
            for job in jobs {
                match job {
                    PendingServerJob::Job(job) => writer.write_object(BatchObject::Job(job)),
                    PendingServerJob::ObjectJob(target, job) => {
                        writer.write_server_object(Some(target));
                        writer.write_object(BatchObject::ObjectJob(job));
                    }
                }
            }
            writer.write_marker(end_marker);
        }
    }

    fn enqueue_batch(&self, batch: CommittedBatch) {
        self.batches.enqueue(batch);
        self.render_loop.wakeup();
    }

    /// Disposes a server object in a batch of its own, outside of the
    /// regular commit.
    ///
    /// The object is disposed, not released: it stays in the table of the
    /// server under its id until the UI-thread object that names it is
    /// dropped, as after an ordinary disposal. An object whose creation has
    /// not been committed yet is created by the same batch, before it is
    /// disposed: upstream the server object exists from the moment the
    /// UI-thread object does, so the disposal always finds it. Left to the
    /// next commit, the creation would follow the disposal and the object
    /// would never be disposed.
    pub(crate) fn oob_dispose(&self, server: ServerObjectId) -> Arc<CompositionBatch> {
        // A disposal that was queued for the next batch is done here.
        self.dispose_on_next_batch.borrow_mut().retain_mut(|entry| {
            if entry.id == server {
                entry.dispose = false;
                entry.release
            } else {
                true
            }
        });
        let pending_creation = {
            let mut creations = self.pending_creations.borrow_mut();
            creations.iter().position(|(id, _)| *id == server).map(|index| creations.remove(index))
        };
        let batch = CompositionBatch::new();
        let mut changes = self.batches.rent_data();
        {
            let mut writer = BatchStreamWriter::new(&mut changes);
            if let Some((id, factory)) = pending_creation {
                writer.write_marker(BatchMarker::CreateStart);
                writer.write(1i32);
                writer.write_server_object(Some(id));
                writer.write_object(BatchObject::Create(factory));
            }
            writer.write_marker(BatchMarker::RenderThreadDisposeStart);
            writer.write(1i32);
            writer.write_server_object(Some(server));
            writer.write(true);
            writer.write(false);
        }
        self.enqueue_batch(CommittedBatch { batch: batch.clone(), changes, committed_at: self.clock_elapsed() });
        batch
    }

    /// Allocates the id of a new server object and queues its creation for
    /// the next batch. `factory` runs on the render thread.
    pub fn create_server_object(
        &self,
        factory: impl FnOnce(&Rc<ServerCompositor>, ServerObjectId) -> Rc<dyn IServerObject> + Send + 'static,
    ) -> ServerObjectId {
        self.dispatcher.verify_access();
        let id = match self.free_server_object_ids.borrow_mut().pop() {
            Some(id) => ServerObjectId(id),
            None => {
                let id = self.next_server_object_id.get();
                self.next_server_object_id.set(id + 1);
                self.server_object_id_in_use.borrow_mut().push(false);
                ServerObjectId(id)
            }
        };
        self.server_object_id_in_use.borrow_mut()[id.index()] = true;
        self.pending_creations.borrow_mut().push((id, Box::new(factory)));
        self.request_commit_async();
        id
    }

    /// Queues an object whose changes must be written to the next batch.
    pub fn register_for_serialization(&self, composition_object: Rc<dyn ICompositorSerializable>) {
        self.dispatcher.verify_access();
        let key = composition_object.serialization_key();
        if self.object_serialization_hash_set.borrow_mut().insert(key) {
            self.object_serialization_queue.borrow_mut().push_back(composition_object);
        }
        self.request_commit_async();
    }

    /// Disposes a server object with the next batch, and releases it: the
    /// server compositor drops it and its id can be reused.
    pub fn dispose_on_next_batch(&self, obj: ServerObjectId) {
        if self.queue_disposal(obj, true, true) {
            self.request_commit_async();
        }
    }

    /// Disposes a server object with the next batch but keeps it, under its
    /// id, until [`release_with_a_later_batch`](Self::release_with_a_later_batch):
    /// the disposal of a composition object, whose server object stays
    /// reachable by the jobs and animations that refer to it while the
    /// UI-thread object is alive, as upstream the UI-thread object holds its
    /// server object.
    pub(crate) fn dispose_and_keep_on_next_batch(&self, obj: ServerObjectId) {
        if self.queue_disposal(obj, true, false) {
            self.request_commit_async();
        }
    }

    /// Disposes a server object with a later batch, without asking for a
    /// commit: used when a UI-thread object is dropped, which must not
    /// schedule work by itself.
    pub(crate) fn dispose_with_a_later_batch(&self, obj: ServerObjectId) {
        self.queue_disposal(obj, true, true);
    }

    /// Releases a server object that was already disposed with a later
    /// batch, without asking for a commit: used when a disposed UI-thread
    /// object is dropped.
    pub(crate) fn release_with_a_later_batch(&self, obj: ServerObjectId) {
        self.queue_disposal(obj, false, true);
    }

    /// Adds `obj` to the disposal list of the next batch, or merges the
    /// request into its entry. Returns whether the list changed.
    fn queue_disposal(&self, obj: ServerObjectId, dispose: bool, release: bool) -> bool {
        // An id that was released already names no object: asking again
        // for its release must not return it to the free list twice.
        if !self.server_object_id_in_use.borrow().get(obj.index()).copied().unwrap_or(false) {
            return false;
        }
        let Ok(mut list) = self.dispose_on_next_batch.try_borrow_mut() else { return false };
        match list.iter_mut().find(|entry| entry.id == obj) {
            Some(entry) => {
                let changed = (dispose && !entry.dispose) || (release && !entry.release);
                entry.dispose |= dispose;
                entry.release |= release;
                changed
            }
            None => {
                list.push(PendingDisposal { id: obj, dispose, release });
                true
            }
        }
    }

    /// Enqueues a callback to be run before the next commit.
    pub fn request_composition_update(&self, action: impl FnOnce() + 'static) {
        self.dispatcher.verify_access();
        self.invoke_before_commit_write.borrow_mut().push_back(Box::new(action));
        self.request_commit_async();
    }

    /// Posts a job that runs on the render thread with the next batch:
    /// before the composition targets are rendered, or after them when
    /// `post_target` is set.
    pub fn post_server_job(&self, job: impl FnOnce(&ServerCompositor) + Send + 'static, post_target: bool) {
        self.dispatcher.verify_access();
        let list = if post_target {
            &self.pending_server_compositor_post_target_jobs
        } else {
            &self.pending_server_compositor_jobs
        };
        list.borrow_mut().push(PendingServerJob::Job(Box::new(job)));
        self.request_commit_async();
    }

    /// Posts a job that runs on the render thread with the next batch and
    /// receives the server object `target` (`None` when it does not
    /// exist). Upstream a job holds the server object itself; this job
    /// names it, and the object is resolved when the batch is read, so the
    /// job reaches an object that the same batch disposes.
    pub fn post_server_object_job(
        &self,
        target: ServerObjectId,
        job: impl FnOnce(&ServerCompositor, Option<Rc<dyn IServerObject>>) + Send + 'static,
        post_target: bool,
    ) {
        self.dispatcher.verify_access();
        let list = if post_target {
            &self.pending_server_compositor_post_target_jobs
        } else {
            &self.pending_server_compositor_jobs
        };
        list.borrow_mut().push(PendingServerJob::ObjectJob(target, Box::new(job)));
        self.request_commit_async();
    }

    /// [`invoke_server_job_async`](Self::invoke_server_job_async) with a job
    /// that receives the server object `target`; see
    /// [`post_server_object_job`](Self::post_server_object_job).
    pub fn invoke_server_object_job_async<T: Send + 'static>(
        &self,
        target: ServerObjectId,
        job: impl FnOnce(&ServerCompositor, Option<Rc<dyn IServerObject>>) -> Result<T, ServerJobError>
            + Send
            + 'static,
        post_target: bool,
    ) -> ServerJobTask<T> {
        let task = ServerJobTask::new();
        let completion = task.clone();
        self.post_server_object_job(
            target,
            move |compositor, object| match job(compositor, object) {
                Ok(result) => completion.set_result(result),
                Err(error) => completion.set_exception(error),
            },
            post_target,
        );
        task
    }

    /// Runs `job` on the render thread with the next batch and returns a
    /// task that completes with its result (`InvokeServerJobAsync`). A job
    /// that fails faults the task with its error.
    pub fn invoke_server_job_async<T: Send + 'static>(
        &self,
        job: impl FnOnce(&ServerCompositor) -> Result<T, ServerJobError> + Send + 'static,
        post_target: bool,
    ) -> ServerJobTask<T> {
        let task = ServerJobTask::new();
        let completion = task.clone();
        self.post_server_job(
            move |compositor| match job(compositor) {
                Ok(result) => completion.set_result(result),
                Err(error) => completion.set_exception(error),
            },
            post_target,
        );
        task
    }

    /// Renders a visual and its children into a new bitmap, on the render
    /// thread with the next batch, after the composition targets are
    /// rendered (`CreateCompositionVisualSnapshot`).
    ///
    /// The task is faulted with a [`CompositionVisualSnapshotError`] when
    /// the visual belongs to another compositor or is not attached to a
    /// composition target (`InvalidOperationException` upstream, which the
    /// asynchronous method also reports through its task).
    pub fn create_composition_visual_snapshot(&self, visual: &CompositionVisual, scaling: f64) -> ServerJobTask<Bitmap> {
        if !Rc::ptr_eq(visual.compositor(), &self.this()) {
            return ServerJobTask::from_exception(Arc::new(CompositionVisualSnapshotError(
                "the visual belongs to another compositor",
            )));
        }
        if visual.root().is_none() {
            return ServerJobTask::from_exception(Arc::new(CompositionVisualSnapshotError(
                "the visual is not attached to a composition target",
            )));
        }
        // The job names the server visual by id, as the other jobs of an object do: the id
        // stays bound to the server visual while the visual is alive.
        let server = visual.server();
        self.invoke_server_job_async(
            move |compositor| match compositor.get::<ServerCompositionVisual>(server) {
                Some(visual) => Ok(Bitmap::from_impl(compositor.create_composition_visual_snapshot(&visual, scaling, true))),
                None => {
                    let error: ServerJobError =
                        Arc::new(CompositionVisualSnapshotError("the server visual no longer exists"));
                    Err(error)
                }
            },
            true,
        )
    }

    /// `GetRenderInterfacePublicFeatures`: the public features of the
    /// render interface, handed to `f` inside the compositor lock; what `f`
    /// takes from them is the outcome.
    ///
    /// As upstream: the features that are cached are the answer at once; a
    /// compositor whose loop does not run in the background asks the server
    /// on this thread (which creates the backend context when needed); a
    /// compositor confined to its render thread without a cache asks with a
    /// job of the render thread (`InvokeServerJobAsync(Server.RT_GetRenderInterfaceFeatures)`)
    /// and the outcome comes with the job, so the caller awaits it. `f` then
    /// runs in the job, on the render thread.
    ///
    /// Upstream hands the dictionary itself to the caller, and the caller
    /// takes the feature from it; here a feature is an `Rc` of the server
    /// side, so `f` takes what it needs inside the lock, where the features
    /// are, and only what it returns leaves. Its third argument tells
    /// whether the calling thread may create the backend context: `false`
    /// only on the thread of a confined compositor, which answers from the
    /// cache. Without a backend context a compositor that is not confined
    /// and whose render interface is not ready has no features (`None`).
    fn get_render_interface_public_features<R: Send + 'static>(
        &self,
        f: impl FnOnce(&ServerCompositor, &HashMap<std::any::TypeId, Rc<dyn std::any::Any>>, bool) -> R + Send + 'static,
    ) -> LocalFuture<Option<R>> {
        self.dispatcher.verify_access();
        if self.is_confined_to_render_thread() {
            // The features are objects of the server side: they are read
            // under the lock, and `f` runs there; it is only taken out of
            // its cell when there is a cache to give it.
            let f = Cell::new(Some(f));
            let cached = self.server.with(|server| {
                let features = server.at_try_get_cached_render_interface_features()?;
                let f = f.take().expect("taken once");
                Some(f(server, &features, false))
            });
            if let Some(result) = cached {
                return Box::pin(std::future::ready(Some(result)));
            }
            let f = f.take().expect("taken once");
            let task = self.invoke_server_job_async(
                move |server| {
                    let features = server.rt_get_render_interface_features();
                    Ok(f(server, &features, true))
                },
                false,
            );
            return Box::pin(ServerJobResult::new(task));
        }
        let result = self.server.with(|server| {
            if let Some(features) = server.at_try_get_cached_render_interface_features() {
                return Some(f(server, &features, true));
            }
            if !server.render_interface().is_ready() {
                return None;
            }
            let features = server.rt_get_render_interface_features();
            Some(f(server, &features, true))
        });
        Box::pin(std::future::ready(result))
    }

    /// Attempts to query for a feature from the platform render interface
    /// (`TryGetRenderInterfaceFeature`): a future that completes with the
    /// feature, or `None` when the render interface has none. Awaited on the
    /// thread of the compositor; see
    /// [`get_render_interface_public_features`](Self::get_render_interface_public_features)
    /// for when it completes.
    ///
    /// Upstream hands out the feature object itself, which the two threads
    /// then share. Here the feature is an `Rc` of the server side, so what
    /// is handed out is a [`RenderInterfaceFeature`](super::RenderInterfaceFeature):
    /// a handle to the same feature that stays inside the compositor lock
    /// and lends the feature there. It keeps the feature alive as the
    /// object upstream does.
    pub fn try_get_render_interface_feature(
        &self,
        feature_type: std::any::TypeId,
    ) -> LocalFuture<Option<super::RenderInterfaceFeature>> {
        // The handle is cloned from the map and bound to the lock inside
        // the lock: no count of the server side is touched outside it.
        let locked = self.server.clone();
        let features = self.get_render_interface_public_features(move |server, features, _| {
            let feature = features.get(&feature_type)?;
            Some(super::RenderInterfaceFeature::new(super::server::LockBound::new(&locked, server, feature.clone())))
        });
        Box::pin(async move { features.await.flatten() })
    }

    /// Attempts to query for the GPU interop feature from the platform
    /// render interface (`TryGetCompositionGpuInterop`): a future that
    /// completes with the interop with GPU objects created outside of the
    /// framework, or `None` when the render interface does not support it.
    /// Awaited on the thread of the compositor.
    ///
    /// Upstream awaits the feature (`TryGetRenderInterfaceFeature`) and
    /// constructs the interop where the await continues: on the render
    /// thread, when a job of the render thread gave the feature. Here the
    /// feature is an object of the server side: the server part of the
    /// interop is made from it inside the lock, where the feature is read
    /// (in that job, for a confined compositor without a cache), and this
    /// thread makes the interop around it.
    pub fn try_get_composition_gpu_interop(&self) -> LocalFuture<Option<Rc<dyn super::ICompositionGpuInterop>>> {
        let feature_type = std::any::TypeId::of::<dyn crate::platform::IExternalObjectsRenderInterfaceContextFeature>();
        let locked = self.server.clone();
        let parts = self.get_render_interface_public_features(move |server, features, may_create_context| {
            let external_objects = features
                .get(&feature_type)?
                .downcast_ref::<Rc<dyn crate::platform::IExternalObjectsRenderInterfaceContextFeature>>()?
                .clone();
            // `None` where this thread may not create the context of the
            // render interface and the render thread has none at present.
            super::CompositionInterop::server_part(&locked, server, external_objects, may_create_context)
        });
        let this = self.this();
        Box::pin(async move {
            let part = parts.await.flatten()?;
            let interop: Rc<dyn super::ICompositionGpuInterop> = super::CompositionInterop::from_server_part(&this, part);
            Some(interop)
        })
    }

    /// Whether an object is queued for serialization (for unit tests).
    pub fn unit_test_is_registered_for_serialization(&self, serializable: &dyn ICompositorSerializable) -> bool {
        self.object_serialization_hash_set.borrow().contains(&serializable.serialization_key())
    }

    /// The compositor registered in the service locator, if any.
    pub fn try_get_default_compositor() -> Option<Rc<Compositor>> {
        FerroLocator::current().get_service::<Compositor>()
    }
}

impl Drop for Compositor {
    fn drop(&mut self) {
        match &self.confined_loop_task {
            Some(task) => {
                // The graph holds what only the render thread may drop (the
                // render targets, the graphics context), so its release is
                // the last job of that thread: the task stays in the loop
                // for one more tick, releases the graph there and then has
                // itself removed. Asked for before the lock is entered: no
                // frame starts after this thread has been inside.
                task.request_release();
                // What the graph shares with this thread is dropped here,
                // for the same reason the other modes release all of it
                // here: the count of the handle of the render interface
                // belongs to this thread. (The platform graphics and their
                // ready state are shared by their contracts and go with it.)
                self.server.try_with(|server| server.render_interface().release_platform_handles());
                // A loop that does not tick again keeps the task, and the
                // graph with it: whichever thread drops the loop then leaks
                // the graph instead of dropping it (`LockedServerCompositor`).
                self.render_loop.wakeup();
            }
            None => {
                self.render_loop.remove(&self.loop_task);
                // The server compositor is released here, on the thread of
                // the compositor: a tick in progress on the render thread
                // may still hold the task, and with it the lock object, but
                // not the graph.
                self.server.release();
            }
        }
        // No frame of this compositor applies a batch from here on. The
        // batches that are still in the queue are completed as they are:
        // the media context commits no compositor while one of the batches
        // it committed is not processed.
        self.batches.complete_all();
        let key = self.key;
        // The registry may already be gone when the thread is exiting.
        let _ = COMPOSITORS.try_with(|compositors| {
            if let Ok(mut compositors) = compositors.try_borrow_mut() {
                compositors.remove(&key);
            }
        });
    }
}

/// A server object listed for disposal by the next batch: the server
/// object is disposed when `dispose` is set, and dropped by the server
/// compositor, its id becoming free, when `release` is set.
struct PendingDisposal {
    id: ServerObjectId,
    dispose: bool,
    release: bool,
}

/// The error of a snapshot that cannot be taken (`InvalidOperationException`):
/// see [`Compositor::create_composition_visual_snapshot`].
#[derive(Clone, Debug)]
pub struct CompositionVisualSnapshotError(&'static str);

impl std::fmt::Display for CompositionVisualSnapshotError {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        f.write_str(self.0)
    }
}

impl std::error::Error for CompositionVisualSnapshotError {}

/// A job waiting for the next batch.
enum PendingServerJob {
    Job(ServerJob),
    ObjectJob(ServerObjectId, ServerObjectJob),
}

/// The error of a job run on the render thread: it crosses to the thread
/// that asked for the job.
pub type ServerJobError = Arc<dyn std::error::Error + Send + Sync>;

/// The outcome of a job run on the render thread: what
/// [`Compositor::invoke_server_job_async`] returns (a task upstream).
///
/// The handle is shared by its clones; the job completes it once, from the
/// render thread. Continuations belong to the thread that created the task
/// and run there: right away when the job completes on that thread,
/// otherwise from its dispatcher.
pub struct ServerJobTask<T> {
    shared: Arc<ServerJobTaskShared<T>>,
}

struct ServerJobTaskShared<T> {
    outcome: std::sync::Mutex<ServerJobOutcome<T>>,
    continuations: crate::utilities::ThreadBound<RefCell<Vec<Box<dyn FnOnce()>>>>,
    thread: std::thread::ThreadId,
}

enum ServerJobOutcome<T> {
    Running,
    RanToCompletion(Option<T>),
    Faulted(ServerJobError),
}

impl<T> Clone for ServerJobTask<T> {
    fn clone(&self) -> Self {
        Self { shared: self.shared.clone() }
    }
}

impl<T: Send + 'static> ServerJobTask<T> {
    fn with_outcome(outcome: ServerJobOutcome<T>) -> Self {
        Self {
            shared: Arc::new(ServerJobTaskShared {
                outcome: std::sync::Mutex::new(outcome),
                continuations: crate::utilities::ThreadBound::new(RefCell::new(Vec::new())),
                thread: std::thread::current().id(),
            }),
        }
    }

    pub(crate) fn new() -> Self {
        Self::with_outcome(ServerJobOutcome::Running)
    }

    /// A task that has completed with `result`.
    pub fn from_result(result: T) -> Self {
        Self::with_outcome(ServerJobOutcome::RanToCompletion(Some(result)))
    }

    /// A task that has failed with `error`.
    pub fn from_exception(error: ServerJobError) -> Self {
        Self::with_outcome(ServerJobOutcome::Faulted(error))
    }

    fn run_continuations(shared: &ServerJobTaskShared<T>) {
        let continuations = std::mem::take(&mut *shared.continuations.get().borrow_mut());
        for continuation in continuations {
            continuation();
        }
    }

    fn complete(&self, outcome: ServerJobOutcome<T>) {
        {
            let mut current = self.shared.outcome.lock().unwrap();
            if !matches!(*current, ServerJobOutcome::Running) {
                return;
            }
            *current = outcome;
        }

        if self.shared.continuations.is_on_thread() {
            Self::run_continuations(&self.shared);
        } else if let Some(dispatcher) = Dispatcher::from_thread(self.shared.thread) {
            // Completed by the render thread: the continuations run on the
            // thread of the task.
            let shared = self.shared.clone();
            dispatcher.post(move || Self::run_continuations(&shared), DispatcherPriority::SEND);
        }
    }

    pub(crate) fn set_result(&self, result: T) {
        self.complete(ServerJobOutcome::RanToCompletion(Some(result)))
    }

    pub(crate) fn set_exception(&self, error: ServerJobError) {
        self.complete(ServerJobOutcome::Faulted(error))
    }

    /// Whether the job has run.
    pub fn is_completed(&self) -> bool {
        !matches!(*self.shared.outcome.lock().unwrap(), ServerJobOutcome::Running)
    }

    /// Whether the job has run without an error.
    pub fn is_completed_successfully(&self) -> bool {
        matches!(*self.shared.outcome.lock().unwrap(), ServerJobOutcome::RanToCompletion(_))
    }

    /// Whether the job failed.
    pub fn is_faulted(&self) -> bool {
        matches!(*self.shared.outcome.lock().unwrap(), ServerJobOutcome::Faulted(_))
    }

    /// The error of a failed job.
    pub fn exception(&self) -> Option<ServerJobError> {
        match &*self.shared.outcome.lock().unwrap() {
            ServerJobOutcome::Faulted(error) => Some(error.clone()),
            _ => None,
        }
    }

    /// Takes the result of a job that ran to completion: `None` while it
    /// runs, and once taken; the error of a failed job.
    pub fn take_result(&self) -> Option<Result<T, ServerJobError>> {
        match &mut *self.shared.outcome.lock().unwrap() {
            ServerJobOutcome::Running => None,
            ServerJobOutcome::RanToCompletion(result) => result.take().map(Ok),
            ServerJobOutcome::Faulted(error) => Some(Err(error.clone())),
        }
    }

    /// Runs `continuation` once the job has run; right away if it has.
    /// Called on the thread that created the task.
    pub fn on_completed(&self, continuation: impl FnOnce() + 'static) {
        {
            // The outcome is read under its lock, so a job completing on the
            // render thread either sees the continuation or has completed.
            let outcome = self.shared.outcome.lock().unwrap();
            if matches!(*outcome, ServerJobOutcome::Running) {
                self.shared.continuations.get().borrow_mut().push(Box::new(continuation));
                return;
            }
        }
        continuation();
    }
}

/// A future of the thread of the compositor: what its asynchronous members
/// return (a `ValueTask` upstream), awaited on that thread.
pub type LocalFuture<T> = std::pin::Pin<Box<dyn std::future::Future<Output = T>>>;

/// `await` of the task of a job of the render thread: the result of the
/// job. A job that fails ends the awaiting method with its error, as the
/// exception that `await` rethrows upstream.
struct ServerJobResult<T> {
    task: ServerJobTask<T>,
    registered: bool,
}

impl<T: Send + 'static> ServerJobResult<T> {
    fn new(task: ServerJobTask<T>) -> Self {
        Self { task, registered: false }
    }
}

impl<T: Send + 'static> std::future::Future for ServerJobResult<T> {
    type Output = Option<T>;

    fn poll(mut self: std::pin::Pin<&mut Self>, cx: &mut std::task::Context<'_>) -> std::task::Poll<Self::Output> {
        match self.task.take_result() {
            Some(Ok(result)) => return std::task::Poll::Ready(Some(result)),
            Some(Err(error)) => panic!("{error}"),
            None => {}
        }
        if !self.registered {
            self.registered = true;
            let waker = cx.waker().clone();
            self.task.on_completed(move || waker.wake());
        }
        std::task::Poll::Pending
    }
}