| Age | Commit message (Collapse) | Author | Files | Lines |
|---|
|
Phase 2 of the GPU-rendering plan deliberately targeted a single head
(GpuContext::output: Option<(crtc::Handle, GpuOutput)>) as a narrow
proof that GBM+EGL+DrmCompositor rendering works at all on this
hardware. DrmOutputManager already supports driving several crtcs at
once - initialize_output is a per-crtc call on one shared manager,
the same way anvil drives multiple outputs - so this was purely an
unexercised restriction, not an architectural limit.
GpuContext::output is now GpuContext::outputs: Vec<(crtc::Handle,
GpuOutput)>, and udev/platform.rs calls initialize_output for every
connected head in its own bring-up loop instead of only the first
after that loop finishes. A head this fails for individually (already
logged, not fatal) still just has no entry and falls back to the
existing legacy Pixman path, unchanged from before.
session.rs's VBlank handler and its VT-switch resume path (which
excludes GPU-driven crtcs from the legacy set_crtc reassert loop, a
different device fd that must never issue mode-set commands against a
crtc DrmOutputManager already owns) both now look a crtc up in the
Vec instead of comparing against a single stored one.
render.rs's own render-loop lookup uses direct field access
(gpu.outputs.iter_mut().find(...)) rather than an equivalent
&mut self method: the borrow checker treats a method call as
borrowing all of GpuContext, including gpu.renderer needed a few
lines later for the same head, where direct field access lets it see
the two borrows are disjoint.
Still gated behind SRDWM_GPU=1 (unset by default) and untested on
real multi-monitor hardware with the flag on - this machine has one
display, so the actual multi-head path itself only gets exercised
whenever it's set on hardware that has more than one.
|
|
Two bugs in how a CSD client's own declared shadow-margin offset
(dwindow.geometry().loc) drives where its content actually renders,
both surfaced by a real Firefox window:
1. A live Firefox window was observed reporting loc = (-10, -10) --
negative, despite sync_geometry's tiled-state hint telling it to
reserve no margin at all. effective_frame_of already clamps this
value to non-negative for its own size calc; the content-position
code (both the masked-content-buffer test and the real content
push) didn't, so a negative margin shifted content the wrong way --
away from the border, not toward it. Clamped to match.
2. Separately, and the actual cause of a later "border isn't over the
window" report on the same Firefox window (this time reporting
loc = (10, 10)): the masked/rounded content buffer's own build step
already renders the client's surface tree shifted by -content_offset
so the buffer's own (0, 0) lands exactly on the real, margin-
excluded content top-left (see rounded_content_buffer's own loc
parameter). Placing that already-compensated buffer on screen at a
*second* content_offset-shifted position double-applied the
correction, landing it content_offset pixels too far up and left of
the border wrapping it. Confirmed live via pixel sampling: Firefox's
own chrome rendered starting 10px above the border's nominal top
edge, fully exposed, square, with no border over it at all.
Split the single `pos` into `content_pos` (unshifted - what the
already-compensated masked buffer uses) and `pos` (content_pos further
shifted by content_offset - kept only for the surface_content_elements
fallback, which renders the client's raw surface tree with no prior
compensation of its own and still needs the shift applied once).
The winit/GLES backend's equivalent path doesn't have this bug: its
rounded_content_element renders the client's live texture directly at
`location` with no separate pre-shifted buffer, so a single
content_offset-adjusted position there was already correct.
|
|
Phase 2 (0274273) deliberately shipped without VT-switch support for
the GPU-driven head, documented as an explicit gap rather than a
silent risk. Before testing it live, wire the real fix instead of
finding out empirically - this already burned three real
reboots getting the *legacy* VT-switch path right, and DrmOutputManager
uses a genuinely different API surface (pause()/activate(), calling
through to DrmDevice's own master-lock acquire/release) than the
manual set_crtc+DPMS reassertion register_session_notifier already
does for legacy heads.
PauseSession now also calls DrmOutputManager::pause() when SRDWM_GPU=1
and a GPU context exists - a separate device/fd from the legacy Card,
so purely additive. ActivateSession calls DrmOutputManager::activate
(false), then deliberately does *not* also force a fresh render for
that head specifically: DrmCompositor::render_frame always issues a
full state commit (atomic or legacy, whichever this device negotiated
- see DrmDevice::is_atomic()), not just a buffer swap, so the
existing data.render_udev_frame() call at the end of this handler
already reasserts mode-set and CRTC-active state together for the GPU
head via render.rs's own GPU branch, the same way it always does.
Also fixed a real conflict Phase 2 introduced: the existing legacy
crtc-reassert loop (explicit set_crtc through the legacy Card/fd) used
to run for every head unconditionally, including one now driven by the
GPU path through a completely different DrmDeviceFd - two separate
fds issuing mode-set commands against the same physical CRTC, exactly
the kind of conflict that produced the worst VT-switch
incidents (EBUSY loops) when it was really one fd racing itself. The
GPU-owned crtc (if any) is now excluded from that loop.
|
|
Extends the SRDWM_GPU=1 opt-in path (Phase 1, fa0c7f1) from a
capability probe into an actual, working GPU render pipeline for
exactly one head, per the plan this was built from
the plan file.
gpu.rs: probe now goes all the way through EGLContext, GlesRenderer,
DrmDevice (real DRM device, separate duped fd from the existing legacy
Card), and DrmOutputManager construction, returning both a GpuContext
and its DrmDeviceNotifier on success. GpuContext::initialize_output
drives one crtc/mode/connector through DrmOutputManager, storing the
resulting GpuOutput for the render loop to find.
Confirmed while reading smithay's own source directly (not assumed):
DrmDevice::new's disable_connectors parameter is not an atomic-vs-
legacy switch - DrmDevice::create_internal tries atomic capability
first and falls back to a Legacy internal variant automatically,
exposed via DrmDevice::is_atomic(), logged here rather than assumed.
platform.rs: probes at startup, calls initialize_output for the first
head only (Phase 2's deliberate scope - see the plan), registers the
DrmDeviceNotifier as its own calloop event source alongside (not
replacing) the existing legacy DRM-fd registration.
render.rs: render_udev_frame's per-head loop checks, before any of the
existing Pixman-specific element-building logic runs, whether this
head's crtc matches the GPU context's initialized output; if so,
renders a plain clear color through render_frame/queue_frame and
continues to the next head, completely bypassing the Pixman path for
that head. Every other head, and this same head whenever the GPU
context or its output is absent, is entirely unaffected.
session.rs: register_gpu_drm_notifier handles DrmEvent::VBlank by
calling frame_submitted() on the matching GpuOutput (required per
queue_frame's own doc comment, or the swapchain runs out of buffers)
and logs DrmEvent::Error without treating it as fatal.
Deliberately out of scope for this phase (documented in the plan):
window content/decorations/cursor on the GPU head (clear color only),
multi-monitor GPU rendering (one head only), and VT-switch pause/
resume for the GPU head specifically (DrmOutputManager's own pause()/
activate() calls are a different API surface from the
existing manual set_crtc+DPMS reassertion, and porting that pairing
correctly needs its own isolated verification pass).
SRDWM_GPU unset (the default) is unaffected: every step above only
runs when it's set to "1", and every failure at any step falls back
to the existing, untouched Pixman path with a logged reason, same
fallback contract Phase 1 already established.
|
|
The udev backend is, by explicit design, 100% software: PixmanRenderer
compositing into legacy KMS dumb buffers. That was a deliberate choice
for portability (dumb buffers work on essentially any DRM driver,
including a VM with no GBM/3D support), not an oversight - but this
machine's real hardware (Intel UHD 620, i915) should support real
GBM+EGL rendering, and the user has asked for a genuine GPU-accelerated
path, GPU preferred with CPU fallback, built as a separate track that
doesn't risk the working software path.
Adds `udev/gpu.rs::probe`: gated behind `SRDWM_GPU=1` (unset by
default - a no-op, zero behavior change for every session that
doesn't set it), attempts GBM device creation on a duped DRM fd, EGL
display/device creation, and a software-rasterizer check, logging
exactly which step failed if any and falling back silently. Wired in
at udev backend startup, right after the DRM fd is opened.
Deliberately does not yet create an EGLContext, a GlesRenderer, or
touch scanout at all. Reading smithay's own reference compositor
(anvil/src/udev.rs) confirmed it wires GBM+EGL rendering together with
atomic-KMS scanout as one unit via DrmCompositor, not as a renderer
swapped into the existing legacy set_crtc/page_flip flip loop this
backend uses today. Adopting DrmCompositor is separate, larger-scoped
work than "swap the renderer" - it replaces the same UdevHead
mode-set/flip machinery the VT-switch fixes
(register_session_notifier's ActivateSession arm, copy_and_flip's
retry backoff) live in, and needs its own plan. This probe answers the
first question - does the hardware even support it at all - safely,
before that larger integration is scoped and attempted.
Cargo.toml: added backend_egl/backend_gbm smithay features, additive
to the existing renderer_pixman path (unchanged, still the default).
|
|
apply_geometry and restore - the Platform callbacks core's toggle_
maximize/apply_snap_zone/restore_window drive for a pure geometry
change - only called sync_geometry, never redraw_decoration_buffer.
The cached border-strip/titlebar bitmaps (self.border_top_decorations,
self.border_bottom_decorations, self.decorations) size themselves from
effective_frame, which can differ from w.geometry alone once a CSD
client's own invisible shadow margin is involved (see that function's
own doc comment) - but nothing here rebuilt them right when this
callback changed w.geometry. The next rebuild only happened whenever
this window's client next committed a frame or some other, unrelated
trigger reached redraw_decoration_buffer, not reliably right away.
Confirmed live: maximizing then restoring a Chrome window left its
border strips sized for the maximized frame while its real content had
already settled back to the smaller restored size, immediately and
permanently until some later trigger happened to catch it up - a
real, visible gap between content and border on the far (east/south)
edges, a different bug from the half-pixel corner seam fixed
separately in blend_corner_pixel.
Both apply_geometry and restore now also call redraw_decoration_buffer
right after sync_geometry, in both the udev and winit backends.
|
|
The udev backend's damage tracking only forced a full repaint (ages =
[0, 0]) on a workspace switch or a VT-switch resume. An ordinary
move/resize/open/close/restack within the same workspace relied
entirely on OutputDamageTracker's own per-element diffing to compute
correct damage for the region a window vacated - which doesn't always
hold: maximizing a window over a second one, then un-maximizing, left
a persistent ghost of the second window's old titlebar/status text
sitting in the vacated corner, unchanged across multiple otherwise-
idle frames.
render_udev_frame now also hashes every visible window's id and rect
each frame and forces the same full-repaint reset whenever that
signature changes - the same "defensive, not a fix for a proven bug
in the diffing itself" reasoning the existing workspace-switch reset
already uses, just triggered by a second, complementary condition.
|
|
The kernel revokes every input device fd across a VT switch away.
libinput has a documented pair of calls for this exact case,
suspend()/resume() (libinput_suspend/libinput_resume), which reopen
every device through the session once it is reactivated. This
codebase never called either one, so after switching back to the
compositor's VT, libinput's device list stayed pointed at fds the
kernel had already revoked - reads on them don't error, they just
silently stop producing events, forever. Rendering, DRM/KMS, and
libseat's own session activation all recovered on their own, which is
what made this look like a display bug rather than an input one; it
took three real forced reboots today, with no visible input from
keyboard or mouse for 30+ minutes after switching back to tty1 each
time, to isolate it as this specific missing call.
register_libinput now returns the Libinput context (a clone of the
one already handed to LibinputInputBackend - it's a reference-counted
handle, not a deep copy, and LibinputInputBackend only exposes an
immutable accessor once it's moved into the calloop event source).
register_session_notifier takes that handle and calls suspend() on
PauseSession, resume() on ActivateSession.
|
|
VT-switch resume: reasserting the CRTC's mode-setting state was never
enough on its own - display power (DPMS) is separate KMS state, and
nothing here ever touched it after a real switch-away-and-back. Page
flips kept succeeding with zero errors logged for the rest of a real
30+ minute session while the panel itself simply stayed dark, which is
what actually explains a user report of the screen and keyboard input
never recovering after one VT switch. Sets DPMS-on unconditionally on
every resume now, the same property zwlr_output_power_v1 already
writes for an explicit client request.
Corner rendering: both backends' side-strip crop (the fix that keeps a
window's flat left/right border strips from poking a solid-coloured
square through the top/bottom strip's own rounded curve) only ever
activated for `w.decorated` windows. An undecorated/CSD window's crop
depends on its *content* actually getting masked to match - the
winit backend never checked that at all, so every CSD window with a
nonzero border_width got the uncropped, "staircase" artifact
unconditionally, confirmed live via a highlighted border colour and
raw pixel sampling. Ported the udev backend's own `border_curve_is_safe`
check (decorated OR content-will-be-masked) into winit, using the
cheap "does this surface have subsurface children" test both content-
masking code paths already gate success on, rather than duplicating
either one's real (comparatively expensive) rendering work just to
probe it.
Full workspace build + clippy -D warnings + test suite (378 tests) green.
|
|
The top-rounds-but-bottom-doesn't investigation these were tracking is
closed: live-verified via a real screenshot (pixel-level, not
eyeballed) that both corners round correctly on both a decorated
window and an undecorated/CSD one relying on content masking. Removes
three log::debug! blocks (corner-mask state, TOP/BOTTOM border strip
position dumps, and a raw alpha-byte dump of the bottom border buffer)
that were firing on every single render pass regardless of whether
anything changed, adding real per-frame overhead for output no longer
needed. Build + clippy + test (378 passing) all still green.
|
|
Pre-existing issues in the uncommitted rust-rewrite work (unused
imports, over-arity glyph-drawing functions, a couple of complex inline
types, or_insert_with(T::default) instead of or_default(), and one
unsimplified test-only arithmetic expression), plus two imports left
unused by switching to or_default(). None of these are behavior
changes. Full workspace build + clippy -D warnings + test suite (378
tests) all green after this.
|
|
# Conflicts:
# crates/config/src/engine/general.rs
# crates/wayland/src/udev/drm.rs
# crates/wayland/src/udev/mod.rs
# crates/wayland/src/udev/render.rs
# crates/wayland/src/winit/render.rs
|
|
Fixes a live-reproduced VT-switch busy loop (failed page_flip retried
with no backoff), shadow rendering bleeding onto occluding windows
unclipped, and a winit-backend buffer-age correctness bug that left
stale cross-window pixels on screen. Committing before merging in the
much larger uncommitted rust-rewrite worktree, which independently
touches several of the same files - this is the pre-merge baseline to
diff against, not a claim that these are the final versions of these
fixes.
|
|
Safety commit before reconciling this worktree with main, which has
diverged with its own separate fixes today. Nothing here is reviewed
or curated yet - this exists purely so none of this work can be lost
to a git operation, disk issue, or worktree cleanup while that
reconciliation happens.
|
|
Every pointer motion event re-ran the same popup/layer/content hit-test
from scratch and delivered focus to whatever it found right now - there
was no notion of "a button is held, keep delivering to the surface that
received the press" at all, which is standard, expected Wayland
compositor behavior (every real compositor does this; it's how dragging,
text selection, and scrollbar-thumb dragging all stay coherent even when
the pointer briefly leaves the widget's bounds mid-gesture).
Without it, a real human's hand drifting even slightly outside the
pressed surface mid-drag - trivially easy during a fast, non-perfectly-
straight mouse motion - sent that client an unrequested `leave` event in
the middle of its own gesture. GTK's drag recognizers (a GtkHeaderBar's
move-the-window gesture, concretely) treat a mid-gesture leave as "this
isn't coherent, abort," which reads as "dragging this window by its
title bar does nothing at all" - live-reproduced this work on Nemo,
and consistent with move_request never having fired once all session
despite real attempts.
pointer_button_grab captures the (surface, origin) resolved on a button
press once the held-button count goes from 0 to 1, and every event under
the grab - motion or button, this press's or a later one overlapping
it - is delivered there instead of wherever a fresh hit-test lands,
until every held button is back up.
|
|
sync_layer_visibility could not tell a real hide (null-buffer commit on
an already-visible surface) apart from a layer-shell client's ordinary
realization sequence (commit with no buffer -> configure -> ack-commit
with no buffer again -> attach real content): both look like "committed,
no buffer" from has_buffer alone. Every layer surface's first realization
was spuriously unmapped and immediately remapped, doubling LayerMap
arrange() passes on every single popup open.
Live-reproduced via an AGS peer session: a full-monitor click-outside-to-
close popup surface came back from a hit-test with geometry wider than
the real output after several open/close cycles on a wl_surface GTK had
reused across role destroy/recreate, and sat in the Top layer above every
real window with no input region set - silently swallowing clicks meant
for windows, dropdowns, and CSD title bars alike.
layer_surfaces_shown_once now gates the hide path on a surface having
actually shown a buffer at least once, and is cleared in layer_destroyed
so a reused wl_surface's next role starts clean rather than inheriting
the previous role's flag.
|
|
srdwm never read xdg_surface.set_window_geometry anywhere. A CSD
client (GTK4/Firefox) declares its real visible content as a sub-rect
inset within a larger buffer that also reserves an invisible shadow
margin - that margin stays reserved in the buffer even once the
tiled-state hint tells the client to stop drawing the shadow itself.
Every render path was positioning content at the client's raw buffer
origin instead of subtracting that declared offset, leaving the
margin's width/height as a gap with wallpaper visible through it at
the window's top-left corner. Confirmed by pixel-diffing the gap
against the real wallpaper at that exact screen position: an exact
match, ruling out "just the client's own dark theme."
Fixed in three places that have to move together: udev/render.rs and
winit/render.rs's content-positioning code, and sync_geometry's
space.map_element call. That last one matters as much as the other
two - it's what smithay's Space (and therefore click hit-testing)
reads, so a render-only fix would have traded a visible gap for an
invisible, same-size hit-test offset in the other direction. Also
applied to the new capture-workspace off-screen render for the same
reason.
This likely also explains real dropdown/context-menu misplacement (and
the clicks landing on the wrong spot) in CSD apps: popup positioning
anchors against the same window position this fix corrects.
|
|
resume, plumbing
Bundles the remaining wayland-crate changes built up here,
touching both backends (udev and winit) and the shared input/rendering
code:
- udev/capture.rs: off-screen Pixman render of an arbitrary (not
necessarily on-screen) workspace's window content to a PPM file --
what crates/core's capture-request queue drives, for a workspace
switcher's thumbnail previews. wlr-screencopy structurally can't do
this (it can only see what an output is presenting), which is why
this exists as a separate render path rather than reusing it.
- input::focus_window now also raises the window in smithay's own
Space, not just core's stacking order - Space is what actually
renders on top and what pointer hit-testing reads, so any focus path
that skipped this (an IPC "focus" dispatch, concretely) left a
window genuinely focused while still rendering, and receiving
clicks, underneath whatever was already topmost. Both backends'
poll loops now re-sync this after any IPC mutation.
- udev/session.rs's VT-switch resume fix (drains a stale pending page
flip before reasserting CRTCs) already has its own earlier, cleanly
isolated commit - not duplicated here.
- Assorted decoration/cursor/rounded-corners/output-management/
screencopy/XWayland changes and their cross-backend wiring.
Coarser than the repo's usual one-purpose-per-commit convention,
deliberately - see the core-crate sweep commit's own message for why.
|
|
A flip issued right before a VT switch away could still be undelivered
when the session resumed - the kernel refuses a new page flip on a CRTC
with one already outstanding, which showed up live as a black screen
that never recovered across two switch attempts, with a rapid repeating
"Device or resource busy" loop in the log. Drain and apply any pending
DRM events before reasserting CRTCs and rendering again on resume, so a
stale flip from before the switch can't collide with the fresh one.
|
|
MISSING.md listed the border frame's bottom/left/right strips as
staying square while the top one (and the titlebar above it) rounds --
"unrelated to client content rounding... not attempted." The left/
right strips genuinely can't participate (border_strips' geometry has
them span only the height between the top and bottom strips, no
corner to round), but the bottom strip is exactly the same shape as
the top one and had no reason left to stay square.
decoration::render_border_bottom mirrors render_border_top exactly
(round_bottom_corners mirrors round_top_corners), cached the same way
in a new border_bottom_decorations map, and drawn in both render loops
via the same all-or-nothing occlusion check the top strip already
uses - pulled out of the left/right strips' per-fragment occlusion
splitting into its own dedicated bitmap path, matching top's existing
trade-off (cropping a rounded bitmap's source rect per fragment is
real extra work for a strip this thin) rather than inventing a new one.
One real bug caught before it shipped: round_bottom_corners' corner-
centre math (height - r - 1) panics on unsigned underflow whenever the
radius clamp lands on the strip's own full height (a real, common case
- a 2px-thick test strip hits it immediately). Fixed by computing the
centre as a signed offset instead, mirroring how the existing dx/dy
distance math already avoids the same class of issue.
|
|
Platform::close only ever called w.toplevel(), which is None for an
XWayland window - closing one (the WM's own close binding, or `srd
dispatch close`) silently did nothing at all, on both udev and winit.
Found live: a leftover untitled fullscreen window wouldn't close via
srd dispatch close even after several seconds, tracing back to this.
Fixed by falling back to X11Surface::close() when there's no xdg
toplevel - it already handles both cases smithay-side (a polite
WM_DELETE_WINDOW for a cooperating client, outright destroy_window for
one that doesn't support it), so no new logic was needed, just calling
it.
|
|
~17 comments across the codebase still pointed at udev.rs/winit.rs/
state.rs by their old flat-file names after those became udev/,
winit/, state/ directories - found while auditing what this work
rushed, since the split verification (function/struct-name diffing,
full test suite) checked structural correctness but never comment
accuracy. Updated each to either the specific new file (e.g. "see
state.rs's REPEAT_DELAY" -> "see state/mod.rs's REPEAT_DELAY",
"udev.rs's monitors()" -> "udev/platform.rs's monitors()") or the bare
module name where the reference was already generic ("the udev/winit
backends", not a specific location).
|
|
Pure reorganization, no behavior change - verified by diffing the
function-name and struct-name sets before/after (both identical) plus
a full cargo test pass. Struct definitions (Card, DrmBuffer, UdevHead,
UdevState) and their small impls stay in mod.rs, since default (crate-
scoped) privacy there is visible to every descendant submodule without
further changes. The rest splits by concern:
- render.rs: the per-frame impl CompState block (render_udev_frame and
the gamma/output-power methods) - still one ~520-line function,
left intact rather than decomposed, given how much of its structure
(the self.udev.as_mut() disjoint-borrow pattern threaded through it)
is deliberate and already documented inline.
- outputs.rs: hotplug reprobe/relayout (impl CompState).
- platform.rs: UdevPlatform's struct/connect logic and its `impl
Platform for UdevPlatform`, previously split apart in the flat file
by ~250 lines of unrelated DRM/session code sitting between them.
- drm.rs: mode/CRTC/framebuffer probing and setup.
- session.rs: libseat/libinput/udev-monitor calloop registration and
the libinput event handler.
A few free functions and one struct (ConnectorProbe, bring_up_head,
probe_connected, pick_crtc, the register_* functions) went from
module-private to pub(crate): called across what are now sibling
submodules, which - unlike a defining module's own descendants --
Rust's privacy model doesn't let see each other's private items.
Matches this crate's existing pub(crate) convention rather than
introducing pub(super), which crates/core's manager/ split used
instead to match *that* crate's plain-private convention.
|