Fixes four bugs behind yattee#956 on Apple TV HD (AppleTV5,3, A8,
tvOS 26.5), where video played audio only with a permanently black
picture, never autoplayed, the clock stayed at 0:00, and navigating
during/after playback could freeze the whole app.
1. A8 GL driver deadlock (device-specific). glTexImage2D uploading
float32 data into a float16 texture hangs in libGLImage's
glgProcessPixelsWithProcessor (a dispatch_group_wait that never
returns), permanently wedging mpv_render_context_render. mpv 0.37
(commit 703f1588) turned on dithering (dither-depth=auto) and
LUT-based scalers by default; on the A8 (no GL_EXT_texture_norm16)
both upload float weights via that fatal conversion, and per-frame
CPU plane uploads (videotoolbox-copy / software decode) hit the same
path. Yattee 1.x shipped mpv 0.36 with these off, which is why the
device worked there. Worked around on AppleTV5,3 only: dither-depth=no,
hwdec=videotoolbox (zero-copy via IOSurface), and bilinear scalers.
Confirmed by two on-device lldb backtraces.
2. Swift cooperative-pool starvation (all platforms). MPVClient's event
loop ran as Task.detached and blocked a cooperative-pool thread for
the client's entire lifetime. With two clients alive (active +
pre-warmed) on a 2-core A8, both pool threads were held and every
await in the app starved: the load pipeline froze at its first
Task.sleep, killing autoplay, Now Playing and progress saving. Moved
the event loop to a dedicated Thread; EAGLContext creation now bridges
through GCD instead of Task.detached.
3. Leaked time-update gate (all platforms). PlayerService dropped all
time updates while loadingVideoID was set; when a load task died
mid-flight the flag leaked and the clock stayed at 0:00 during
playback. Heal the gate on the ready/playing transition and leak-proof
play() with a defer; log the previously silent load cancellation.
4. Orphaned player view (all platforms). Detach the shared render view in
MPVContainerView.deinit when no successor container exists, and skip
framebuffer recreation while the view is windowless, to stop a 100% CPU
SwiftUI trait-update loop when opening a settings detail during playback.
Build 264 stutters for a split second every ~10 seconds. The periodic
playback-stats task fetches ~30 debug properties in one ~190ms block on
mpvQueue, and MPVClient.render() wrapped mpv_render_context_render in
mpvQueue.sync - so the render thread queued behind the fetch and stalled
~190ms, at exactly the stats interval. Build 261 predates the stats
task, which is why it was unaffected.
libmpv's render API is documented thread-safe relative to the client
API, so render-context calls no longer go through mpvQueue:
- New renderContextLock protects renderContext lifecycle vs use;
render(), renderWithDepth(), renderSoftware(), reportSwap(),
shouldRenderUpdateFrame() and hasRenderContext use it
- Render-update callbacks dispatch on a dedicated renderEventQueue so
frame-ready notifications can't queue behind property/command work
- destroyRenderContext takes mpvQueue then the lock, still excluding
in-flight renders during teardown
- The stats task skips its property fetch on iOS/tvOS unless verbose
MPV logging is on (macOS keeps it for the render watchdog)
- performRender slow-frame warnings now break down render vs present
time and report frame-delivery gaps for future triage
Since 1d9c5e3b performRender feeds swap timing to mpv via
mpv_render_context_report_swap, but only after presentRenderbuffer -
which is skipped while PiP is active. Once swap reports stop mid-stream,
display-vdrop's vsync estimation goes stale and mpv drops nearly every
frame (vo drop counter climbing ~19/s on a 24fps stream), turning PiP
into a slideshow.
The frame is genuinely displayed during PiP via the
AVSampleBufferDisplayLayer, so report the swap on that path too.
1d9c5e3b removed the main-thread EAGLContext bind from setupAsync so
the context is never left current on two threads. But libmpv's
mpv_render_context_create requires a current GL context on the calling
thread (it probes glGetString(GL_VERSION)), so render context creation
failed with MPV_ERROR_UNSUPPORTED (-18) and every video played with no
--vo ("No render context set") on real iOS/tvOS devices.
Bind the context around the createRenderContext call and unbind right
after. The bind is scoped to setup, before the display link starts and
before any frame is rendered, so the never-left-current invariant that
protects the A10X fence fix (#947/#949) still holds.
Verified on Apple TV 4K (3rd gen): createRenderContext succeeds and
video frames render.
Claude-Session: https://claude.ai/code/session_01UEXP5f6F4S1zLdY8fdVuLJ
Apple TV 4K 1st gen on tvOS 26 crashes to the home screen with a
"blocked by IOFence" gpuEvent at playback start (#949) and stutters
with a starved demuxer cache (#947). Defensive fixes in the shared
iOS/tvOS OpenGL path:
- Drain in-flight GPU work (glFinish) before destroying the
renderbuffer, so a pending presentRenderbuffer fence can't hold the
CAEAGLLayer IOSurface being torn down or reallocated
- Never leave the EAGLContext current on the main thread: framebuffer
create/destroy and the background glFinish now unbind on exit, and
setup no longer binds the context on main at all
- Disable retained backing on tvOS; it exists only for iOS PiP frame
capture and adds per-frame IOSurface fence pressure
- Report swaps to mpv after each present so display-vdrop (the tvOS
default video-sync) gets swap-timing feedback, matching macOS
Add verbose-gated diagnostics for remote triage: playback stats every
10s (cache fill/rate, dropped frames, hwdec, avsync) and a rate-limited
slow-frame warning in performRender.
Addresses #949 and #947
