On a Wayland HDR session, colors flickered between HDR and washed-out SDR every ~5 seconds: the mpv-leg timeout in apply_hdr_state was armed after SetHdrOutput returned, so a synchronous reply (the no-op short-circuit that runs on every playback restart and re-describe) orphaned the timer. Five seconds later it fired against a healthy plane and withdrew the live HDR image description; the next re-apply re-attached it and armed the next orphan.
Arm the timeout only while the reply is genuinely outstanding. An asynchronous reply still removes it, and a genuinely wedged core is still bounded.
Verified in a headless-sway Docker repro against the same PQ/BT.2020 file: 2.15.0 logs the false 'mpv never answered the output colour-space switch' warning 5 s after plane bring-up and again after source recognition; the fixed build logs none across bring-up, HDR playback, and seeks.
close#2016
On a Wayland session whose compositor hands clients no GPU device
(Cinnamon/Muffin 6.6.3 on Mint 22.3), Mesa falls back to llvmpipe and
libva-wayland's vaInitialize segfaults inside mpv_render_context_create
the moment mpv is given the wl_display - the app silently closes when
playback starts. 2.14 was the first release to hit this because its
bundled libmpv is the first with VAAPI compiled in, and the device init
runs eagerly at render-context creation, so the hardware-decoding
toggle cannot avoid it.
Skip MPV_RENDER_PARAM_WL_DISPLAY when GL_RENDERER names a software
rasterizer: zero-copy Wayland interop does not exist on llvmpipe anyway,
and mpv's hwdec=auto still reaches the GPU through a DRM render node,
which works even on these sessions. Accelerated sessions keep the
zero-copy handoff unchanged. Verified in a container against Muffin and
weston on llvmpipe: the handle is withheld and playback works.
The display-agnostic texture renderer restored by 9cdfe759 is deleted
again, this time for good: it is a second, SDR-only rendering stack
(isolated EGL context on Flutter's display plus EGL-image handoff) kept
alive solely to host sessions that cannot bring up the Wayland plane -
X11/XWayland or a failed plane bootstrap - and it was the source of the
native lifecycle and EGL state-churn fixes of the 9f2e0507 era. Linux
video now requires a Wayland compositor; a session that cannot host the
plane fails initialization with VIDEO_PLANE_UNSUPPORTED instead of
silently rendering through the second stack.
Reverts the restore commit's machinery: mpv_texture.cc/.h and
mpv_gpu_bootstrap.cc/.h deleted, the plugin's texture-registrar,
bootstrap, and waitForVideoReady paths removed, MpvPlayer's texture-mode
render-context API dropped, and the Dart-side renderMode setting, its
settings tile, translation keys, and the Player textureId member
withdrawn. The #1874 HDR diagnostic work is unaffected.
The retry added while chasing the 2.13.0 hwdec regression never fires:
mpv probes hwdec interop lazily at the first decode attempt, and its
failure does not fail mpv_render_context_create, so the deep config never
gets rejected and the 8-bit tier is never reached. The actual regression
was the libmpv build losing the DRM providers, fixed in the previous
commit. Remove the dead retry and its prefer_deep plumbing, and correct
the pre-flight comment that claimed the interop probe runs at context
creation.
Hardware decoding stopped working for Linux users on 2.13.0 (Fedora 44
report): every source decodes in software, and AV1 plays black video with
audio. Two defects in the pinned libmpv build.
First, mpv's meson 'drm' feature silently disabled itself because the CI
builder lacks libdisplay-info, and every VAAPI path that does not depend
on a display server is derived from it: vaapi-copy's standalone render-node
device (the path 2.12.1 worked on) and the GL dmabuf interop for direct
vaapi. With only the Wayland VA provider compiled in, a machine whose
Wayland VA display fails to initialize has no fallback, and vaapi-copy
has an empty provider list - every source lands on software decoding.
Pin -Ddrm=enabled, -Dvaapi-drm=enabled, -Degl=enabled and
-Dvaapi-wayland=enabled, and add libdisplay-info-dev to the CI package
lists, so a missing piece fails the build instead of shipping silent
software decode.
Second, the bundled static FFmpeg has no AV1 software decoder: its native
av1 codec is hardware-accelerated only, so once hwdec fails there is no AV1
path at all - every packet errors, video hits EOF, the plane goes black
while audio keeps playing. Pin dav1d 1.5.4 (both VideoLAN remotes agree on
the tag object and root commit), build it static before ffmpeg, and pass
--enable-libdav1d.
The build-plan stub test now asserts the hwdec feature flags, the dav1d
static build, and ffmpeg's libdav1d. Verified in an ubuntu:24.04 container
with the production flag sets: meson reports drm, vaapi-drm, vaapi-wayland,
egl and dmabuf-interop-gl enabled, and ffmpeg configures CONFIG_LIBDAV1D=yes
with the AV1 VAAPI hwaccel.
close#1874
The first GitHub Actions build of the Linux path failed on six issues the
macOS host could not catch:
- HandleFrameDone is a static handler; CancelFrameAckWatchdog() needed the
explicit self-> qualification.
- handle_texture_ready_result/handle_ready_timeout call
release_video_resources before its definition; forward-declared.
- finish_leg captured self without using it (Werror).
- CanCommandOutputProperties was private; made public for the kUnknown
live-core check.
- MPV_ERROR_UNKNOWN does not exist in libmpv; the timeout now reports
MPV_ERROR_UNSUPPORTED (any non-success serves the latch; the log line
names the reason).
- The hdr-tone-mapping error path referenced the handler's FlValue value
inside an async lambda; an owned std::string copy is captured instead.
- The texture register-failure response used a heap string freed before the
handler responded; use a literal.
2.13.0 deleted the display-agnostic EGL/Flutter-texture renderer and made
the native Wayland plane the only path, hard-rejecting every session that
cannot host one - X11, XWayland (SteamOS Gaming Mode runs native apps
through Gamescope's XWayland), or a plane whose EGL bootstrap failed. This
restores the 2.11.0 texture path from git history as the fallback:
- Re-add mpv_texture.cc/.h and mpv_gpu_bootstrap.cc/.h (2.11.0 verbatim):
an FlTextureGL whose populate renders mpv into an offscreen FBO sampled
by Flutter via an EGL image.
- MpvPlayer gains the texture-mode render-context API (InitRenderContext,
HasRenderContext, GetEglDisplay/Context, Render(w,h,fbo)) beside the
plane's InitRenderContextForSurface/RenderToSurface. The isolated ES 2.0
context on Flutter's display and the X11 display param are exactly the
2.11.0 configuration hardware decode demonstrably worked in.
- initialize now tries the plane first and falls back to the texture path
when it cannot be brought up, returning the texture id (Dart's 2.11.0
'result is int' contract) with waitForVideoReady gating playback until
the GPU bootstrap settles. Both paths share one mpv core, so the
plane/texture decision precedes render-context creation.
- hdr-enabled/hdr-tone-mapping are intercepted in texture mode (no plane,
no HDR); the HDR toggle hides itself via isHDRSupported.
- New Linux setting 'Video rendering mode' (Automatic / Texture) forces
the fallback - the user-visible workaround for plane-only trouble and
for the hwdec interop regression, plus translations in all locales.
SDR only on the fallback, matching 2.11.0; the plane path is unchanged.
Three failure modes in the 2.13.0 Wayland path, three fixes:
- The mpv leg of an HDR transaction had no timeout: the surface watchdog
re-armed while it ran, but the HDR method call stayed unanswered when mpv
never replied, leaving the transaction queue stuck behind a ghost forever.
A 5 s timeout (sharing the surface's horizon) aborts the transition,
withdraws any description, resumes presentation, and answers the request
exactly once via a shared latch; a late mpv reply self-heals through the
stale-token re-apply path.
- The kUnknown quarantine hid the plane for the whole session when mpv
stopped answering - the AV1-transparent report was a plane hidden this
way while sound kept playing. Hiding is now reserved for a core that is
genuinely going away; a live one presents undescribed (sRGB by protocol),
and a playback restart clears the quarantine so one poisoned source
cannot hide every later one. Product decision: visible-wrong beats
invisible.
- hwdec's dmabuf interop probe runs at mpv_render_context_create time
against the plane's fresh EGL display/context, and drivers that fail it
on a deep config silently land every source on software decoding (the
Fedora 44 report; 2.12.1 created the context on Flutter's display). The
prerequisites are now logged explicitly at creation, and a failed render
context is retried once on the 8-bit config tier - the 2.12.1-equivalent
configuration - with HDR off by the depth gate.
Present() arms wl_surface.frame and frame_pending_ is cleared only by the
frame callback. Compositors are entitled to stop acknowledging frames for
occluded or minimized surfaces - wlroots-lineage compositors (Hyprland) do
exactly that - and nothing bounded the wait: one missed callback froze the
plane on its last buffer forever, because every later render bailed on
frame_pending(). That is the 2.13.0 black-video report on Hyprland: the
first commit is the pre-allocated 1x1 (or pre-video black) buffer, fully
occluded by the opaque Flutter surface, and the callback it armed never
arrives.
Two changes, one stall:
- A 500 ms frame-acknowledgement watchdog in Present(): on expiry the dead
callback is withdrawn and a fresh present is asked for, so the plane
re-commits instead of sitting on the latch. A miss budget (5) stops
poking a surface the compositor is still ignoring; a real
acknowledgement resets it.
- The very first present is refused until mpv actually has a frame: the
first forced render happens at setVideoRect time, before anything is
decoded, and committing that empty buffer is exactly the commit an
occluded surface ignores. The sticky plane_needs_render flag keeps the
owed resize refresh pending until content exists, so the first present
still happens at the right size the moment the first frame lands.
The SET_PROPERTY_FAILED error surfaced to Dart carries only mpv's own text
("unsupported format for accessing property"), which names the failure
mode, never the property. Every report of a refused write is therefore a
guessing game. Include name=value in the error description on all three
setProperty branches (generic, hdr-enabled, hdr-tone-mapping) and warn with
the same pair from MpvPlayer::SetPropertyAsync so the HDR transaction's
target-* writes — which never reach the channel — are attributable too.
mpv_request_log_messages is a single global level, and the vaapi probe and
"Using software decoding" fallback are MSGL_INFO, so at "warn" a silently
software-decoding session leaves no trace in the app log. Raise the request
level to "info" and observe hwdec-current natively, logging every decode
path transition from the player instead of relying on an overlay readout.
* fix(linux): defer buffer_scale until the first frame is presented
The video plane's wl_egl_window starts as a 1x1 placeholder. When playback starts, SetRect() sends wl_surface.set_buffer_scale(2) and resizes the window, but mesa commits the EGL surface's pre-allocated 1x1 back buffer on the first eglSwapBuffers. A 1x1 buffer is not an integer multiple of scale 2, so the compositor raises WL_SURFACE_ERROR_INVALID_SIZE, tears down the Wayland connection and plezy exits (issue #1872). At display scale 100% the same 1x1 buffer is legal, which is why that workaround worked.
Defer sending the new scale until after the first commit: the first buffer is 1x1 at scale 1 (always legal), then the scale change lands on the wire and applies to the next commit, whose buffer mesa allocates at the resized window size. The one-frame 1x1 flash is mpv's black first frame - imperceptible.
* fix(linux): gate buffer_scale on a first-frame latch, reset state on destroy
Address review: buffer_attached_ cannot represent "first frame
presented" - DetachBuffer() clears it while the committed scale stays
on the wire, leaving a crash path (present at scale 2, detach, move to
a scale-1 display, then a swap commits a scale-1 buffer while scale 2
is still active).
Replace it with first_frame_presented_, set once a frame has been
presented and cleared only when the wl_surface is torn down, so scale
changes queue even with no buffer attached. Reset scale_sent_ to 1 in
Destroy() too: a freshly created wl_surface starts at scale 1, and a
stale value would suppress the first scale request after recreation.
The plane added three runtime libraries that bundle-libs.sh deliberately does not
bundle, so they have to be declared per distro by hand - and two hand-maintained
lists drifting apart is the failure this guard exists to prevent.
check_linux_package_deps.py parses the runner's CMake for every pkg-config module
it links, follows target_link_libraries to prove each one actually reaches the
binary, and requires a package name for it in every distro's depends list. It
fails closed on the shapes a naive parser gets wrong: a pkg_check_modules call
naming several modules, options preceding the module name, and version
constraints like mpv>=0.40 that would otherwise be read as a package nobody
ships.
The smoke job builds the three packages and reads the dependencies back out of
the artifacts, deriving what to expect from build-packages.py rather than
restating it - so a library is declared once and verified everywhere. That job is
off by default, which is exactly why it must not carry its own copy of the list.
The Linux native job names libwayland-dev and libegl-dev instead of riding
GTK's and epoxy's transitive dev dependencies, matching the CMake comment's own
rationale. In CI the host-dependency guard runs once: the named step covers the
staged bundle, and build-packages.py's internal run - which exists for by-hand
packaging - is skipped. The smoke job also drops patchelf, which nothing
invokes.
Video on Linux went through a Flutter texture: 8-bit sRGB, which cannot carry
HDR at all, and which forced a whole-window Flutter recomposite for every video
frame. This moves it onto a wl_subsurface stacked below the Flutter surface, with
mpv rendering into an EGL window surface on it through the libmpv render API. The
subsurface is desynchronized, so video and UI now present independently.
With the plane in place HDR follows: the surface is described to the compositor
through wp_color_manager_v1 as the source's own curve and gamut - PQ or HLG,
BT.2020 - carrying whatever HDR10 static metadata the stream actually declares.
The description and the buffer it describes land on the same commit, staged and
validated before mpv is switched, so a PQ frame is never presented labelled sRGB.
A five-second watchdog bounds the one wait a compositor could otherwise leave
hanging. A session that cannot host the plane - X11, or a compositor without
wl_subcompositor - fails initialize with VIDEO_PLANE_UNSUPPORTED naming the
reason: the texture path is gone, and refusing by name beats degrading to
something the user cannot see. An SDR output, a missing capability or an 8-bit
config keep the plane and simply leave it undescribed.
The output's colour state is trusted only when it has been earned. Every landed
property step records itself as it lands; a reset or sequence that cannot
finish downgrades its result to unknown and marks the applied-output cache
untrusted until a clean apply earns it back. A plane whose output state cannot
be named is quarantined - hidden, its description withdrawn - and the
quarantine is recorded state: an unrelated visibility change cannot put a
mislabelled plane back on screen, and only a commit that resolves to a nameable
outcome lifts it. A rect collapsing to zero detaches the buffer exactly as
hiding does, a refused setVideoRect drops the Dart-side sent-rect cache so the
next layout pass retries for free, and a refused tone-mapping pick tells the
user instead of dying in a log.
NVIDIA's Wayland EGL (through at least 610.xx) offers no 10-bit unorm window
configs, so the plane takes half-float as the tier between 10-bit unorm and
8-bit, declares the whole surface opaque so the compositor never reads the
alpha those configs carry, and states GL_RGBA16F rather than a 10-bit lie.
Whether the output is in HDR is read from luminance headroom above its own
reference white rather than from the preferred transfer function, which current
KWin no longer answers PQ for; the margin is half a stop, because KWin reports
an undimmed maximum over a software-dimmed SDR white. Validated on an RTX 4090
(driver 610.57.04) under KWin 6.7.4 with locked-exposure photographs.
Who tone-maps is a user choice. The default is the compositor: photographed on a
400-nit HDR output against a PQ chart it keeps 400 -> 1000 nits monotonic and
separated where the player leg flattens them, because the player path drives
mpv's legacy vo_gpu, whose own standalone output scores the same. The gap is the
renderer, not the wiring.
The decision itself - what the source carries, what the output supports, what to
tell mpv and what to tell the compositor - lives in hdr_metadata.h, free of
Wayland and GTK so its luminance validation can be tested without a display
server. Sending an incoherent luminance set is a protocol error that disconnects
the client, so the rules are worth a unit test.
The deb, rpm and pacman packages now declare wayland-client, wayland-egl and EGL:
the plane links them directly and bundle-libs.sh deliberately never bundles them,
since they are coupled to the running compositor and GPU driver.
lib/dev/harness_main.dart is a second entrypoint for measuring this on hardware -
it drives one clip with scripted mpv properties and reports the colour state mpv
actually settled on. Nothing imports it, so it is tree-shaken out of the app.
Verified on a Steam Deck against an external 400-nit HDR display: the compositor
reports PQ / BT.2020, the connector carries HDR_OUTPUT_METADATA, and against mpv
vo=gpu-next on the same frame the shipped build sits 4.90 counts away overall -
closer to the reference HDR player than to its own SDR fallback.
wayland-scanner output for the staging colour-management protocol, which the
native video plane uses to describe itself to the compositor as PQ / BT.2020.
Committed rather than generated at build time: the protocol only appeared in
wayland-protocols 1.41, newer than the version the distributions this app is
built for ship. Vendoring keeps the build working regardless of the host and
adds no build dependency on wayland-scanner. Generated from wayland-protocols
1.49 with wayland-scanner 1.25.0.
Nothing links these yet; the CMake wiring and the plane that uses them follow.
The local .clang-format exempts the generated sources from the runner's style.
This was resulting in two audio stutters per playback resume.
Regression originates in 9f2e05079 (release 2.10.0).
Co-authored-by: Torin Cooper-Bennun <torin.cbennun@googlemail.com>
Every URL the player opens is a media-server stream or a local file, so mpv's
bundled ytdl_hook has nothing to resolve. It still ran an on_load hook per
open and, whenever an open failed, spawned yt-dlp with the full stream URL in
its argv — access token included, readable through /proc on Linux. It also
added ~700ms to every failed open and buried the real "[stream] Failed to
open" line under three ytdl_hook errors.
mpv decides whether to load the builtin script inside mpv_initialize, so this
has to be an option set beforehand rather than a property set from Dart.
Verified against mpv 0.41: --ytdl=yes logs "Loading lua script
@ytdl_hook.lua", --ytdl=no never loads it.
Apple is deliberately excluded: the bundled libmpv is built without Lua, so
the option does not exist there and setting it would only print an mpv error
on every player init.
The Linux native reliability job stopped compiling mpv_player.cc: the
node-conversion builder exposed a leaf named `Bool`, and X11's Xlib.h —
reached through epoxy/egl.h -> EGL/eglplatform.h — defines `Bool` as a
macro for `int`, so the declaration was rewritten into nonsense.
Renames that leaf to `Boolean` across the shared walk and all three
builders. The name is the only thing that changes; no conversion
behaviour differs.
Consolidates duplicated logic behind shared implementations — paginated
grid tabs, focus chrome, cached remote stores, sheet selection columns,
the server artifact store and a test fixture layer — and removes code
that had become unreachable. Net reduction of about 5,500 lines with no
behaviour change.
Where a fix had landed separately in code that moved into a shared
helper, the fix was re-applied inside the helper rather than left behind
in the copy that went away.
Introduces shared seams for paginated views, D-pad reorder, media control
routing, async singletons and the device method channel, then points the
open-coded copies at them.
Also removes unused models and duplicated provider/server plumbing, folds
the twice-implemented artifact store in the server, and factors the
repeated Flutter toolchain prologue in CI into a composite action.
Audio-only mpv core on every platform (dedicated
com.plezy/mpv_audio_player channels): parameterized android/windows/
linux mpv plugins and a new apple MpvAudioPlayerCore, all skipping
video/window paths (vid=no, audio-display=no, gapless-audio=weak).
MusicPlaybackService drives an in-memory queue with shuffle/repeat,
file-loaded-event gapless arming (property edges coalesce and the
android bridge drops them), per-track progress reporting, OS media
controls, audio focus, sleep timer, and error auto-skip.
PlaybackCoordinator enforces one live native player: starting video
disposes the audio core first.
edde746