Point the redox-drm recipe at the new durable fork at
local/sources/redox-drm/. The source tree is now an absolute
symlink from local/recipes/gpu/redox-drm/source to the fork,
matching the pipewire / wireplumber fork model. The absolute
target survives moving the project root.
Drop the dead recipe-local patches (P5, P8, P10 — P9 was dropped
in the prior redox-drm recipe commit). All four were already
applied to the in-tree source before the fork move, so the patch
artefacts were inert. The fork is the durable source of truth;
future Red Bear changes go as git commits there.
Bump version 0.1.0 -> 0.2.3 and add the v6.0 2026 Red Bear fork
marker to the description. The recipe documentation now describes
the supported GPU targets (virtio-gpu, Intel Meteor Lake / Arrow
Lake / Lunar Lake, AMD display glue) and the hardware validation
status of each.
The amdgpu recipe's idr_* / ida_* header conflict with linux-kpi is
now resolved (see the preceding commit). Re-enable amdgpu in
redbear-full.toml by changing amdgpu = "ignore" to amdgpu = {} and
add a [package] section to the amdgpu recipe marking the 0.2.3
release and describing the bounded display-glue compile surface
(DCN20 / DCN30 / DCN31 backend files plus the Rust-side init /
connector detection / modeset glue).
The linux-kpi idr.h header and the amdgpu recipe's redox_glue.h
both define struct idr and the idr_init/alloc/remove/find/destroy
inline functions, but with different members and different argument
signatures. When the amdgpu build includes both headers in the same
translation unit (via -include linux/amdgpu_stubs.h and via
redox_glue.h), the compiler reports a redefinition of struct idr
and conflicting types for the idr_* inlines, blocking the build.
Resolve the conflict by gating the linux-kpi idr declarations on a
new REDBEAR_AMDGPU_BUILD preprocessor flag. The amdgpu recipe's
CFLAGS now defines REDBEAR_AMDGPU_BUILD, so the linux-kpi
declarations are suppressed and redox_glue.h's authoritative copies
take over. Every other consumer of linux-kpi (the redox-drm scheme
daemon, the Wi-Fi drivers) continues to see the generic stubs.
The IDA macros and struct ida in linux-kpi are kept outside the
gate because they are not part of the amdgpu surface and must
remain available to other drivers. A forward declaration of
struct idr is added at the top of the header so the
struct ida { struct idr *idr; } member compiles cleanly when the
IDR definitions are suppressed. The IDA macros are updated to
no-ops against a pointer-typed idr slot; the amdgpu build does
not exercise them and other consumers continue to treat them as
stubs.
Point the redox-drm recipe at the new durable fork at
local/sources/redox-drm/. The source tree is now a symlink from
local/recipes/gpu/redox-drm/source to the fork, matching the
redox-driver-sys / pipewire / wireplumber fork model.
Drop the four dead recipe-local patches (P5, P8, P9, P10). All four
were already applied to the in-tree source before the fork move, so
the patch artefacts were inert. The fork is the durable source of
truth; future Red Bear changes go as git commits there.
Bump version 0.1.0 -> 0.2.3 and add the v6.0 2026 Red Bear fork
marker to the description. The recipe documentation now describes
the supported GPU targets (virtio-gpu, Intel Meteor Lake / Arrow
Lake / Lunar Lake, AMD display glue) and the hardware validation
status of each.
atomic_check previously ignored the _available_connectors parameter
(prevented by underscore prefix). The CRTC state's connectors: Vec<u32>
field declared which connectors to bind, but atomic_check never
verified they actually existed or were connected. This allowed
client commits to reference phantom or disconnected connectors
and silently produce invalid state.
Fix: use the available_connectors slice to validate that each
referenced connector ID exists in hardware and has connection
status Connected. Return CrtcNotFound or ConnectorDisconnected
respectively so the kernel rejects invalid commits with a clear
error rather than producing a malformed display state.
This unblocks libdrm drmModeAtomicCommit callers that rely on
DRM_MODE_ATOMIC_ALLOW_MODESET returning success only for valid
connector configurations.
R12: All 5 ConnectorInfo construction sites (kms/connector.rs
synthetic_displayport, intel/display.rs, intel/mod.rs ..connector
inheritance, virtio/mod.rs both sites, amd/display.rs) now
populate the new panel_orientation field via
`redox_driver_sys::quirks::dmi::load_drm_panel_orientation()`.
The 36 panel orientation rules from 50-drm-panel.toml (Linux
7.1 drm_panel_orientation_quirks.c: Jupiter 0x0B57, Galileo
0x0B47, etc.) now apply to every connector detected.
R13: main.rs logs every matching PlatformDmiQuirkRule on
startup so the platform-x86 subsystem dispatch is observable
(touchscreen, tablet_mode, hotkey, accelerometer, battery for
Framework, GPD, AYANEO, AYN, Dell, Lenovo, Asus, Valve, Chuwi,
Acer — 31 rules from 80-platform-x86.toml).
R21: main.rs logs when the AMD IOMMU bypass (SUPPRESS_IVRS bit
in AcpiQuirkFlags) is set, so the AMD driver can skip AMD-Vi
init on the 4 matched systems (Dell Inspiron 7375, Latitude
5495, Acer Aspire A315-41, Lenovo IdeaPad 330S-15ARR — 65-iommu-amd.toml).
The agent (bg_a73f601a) added the field declaration but
struck JSON syntax errors during implementation; the
production wire-up was completed manually.
cargo check: builds clean (pre-existing E0382 in drivers/fence.rs
test is unrelated).
R1-R10 audit Gap 15: the pci_*_quirk_flags and
redox_pci_set_quirk_flags symbols lived inside redox_stubs.c
alongside kmalloc, printk, and other generic glue functions.
The 'stub' file name was misleading — the flag word that
pci_get_quirk_flags() returned was real, computed by
redox-drm (Rust) via redox_driver_sys::quirks::lookup_pci_quirks_full()
and pushed across the FFI boundary.
This change:
- Adds source/redox_quirk_bridge.c containing the three
symbols plus a static g_redox_quirk_flags global. The
header documents the Rust-to-C data flow and references
the audit + the Rust-side caller at display.rs:155.
- Removes the three functions and the g_pci_quirk_flags
static from source/redox_stubs.c. redox_stubs.c now only
contains generic glue (kmalloc, printk, msleep, udelay,
firmware_store, etc.) and the file name matches its
contents.
- Updates recipe.toml Stage 1 to compile the new
translation unit alongside redox_stubs.c. Both files
are linked into libamdgpu_dc_redox.so.
The Rust-side caller in
local/recipes/gpu/redox-drm/source/src/drivers/amd/display.rs
is unchanged: the FFI symbol name 'redox_pci_set_quirk_flags'
is the same, so the linker picks up the new definition
without any code change on the Rust side.
No caller code in amdgpu_redox_main.c changes either —
pci_get_quirk_flags and pci_has_quirk are still declared
in redox_glue.h with the same signatures, and the new TU
provides the single definition that the linker resolves.
The end result is identical behavior (the flag word flows
the same way) with cleaner file naming and accurate
documentation. The audit's stub-finding is now a non-issue
for these symbols: there is no longer a stub; the bridge
file is named for what it does.
Replace stub ioctl handlers with real per-fd magic token assignment
and master tracking. First card opener auto-becomes master. Master
state is tracked and cleared on fd close.
When pcid-spawner launches redox-drm for a GPU device, the init
service (10_redox-drm.service) may also start a second instance.
The guard check in the service file (head -c 1 /scheme/drm/card0)
has a race condition with pcid-spawner's scheme registration.
Move the scheme:drm existence check into the binary itself. If
scheme:drm is already registered when run() starts, log and exit
gracefully with daemon.ready() instead of crashing with a fatal
"no GPU found" error.
Critical fixes from Linux kernel cross-reference:
- execlists.rs flush_pending(): write upper 32 bits FIRST, then lower.
BSpec requires upper-first on legacy ELSP; GPU latches on lower write.
Previously wrote lower then upper — stale upper dword used by GPU.
- execlists.rs init(): write only INHIBIT_SYN_CTX_SWITCH to RING_CONTEXT_CONTROL.
Linux init_common_regs() disables ENGINE_CTX_RESTORE_INHIBIT and
RS_CTX_ENABLE for normal operation. Our old code set restore inhibit,
preventing context save/restore on context switch — GPU would hang.
- context.rs: LRC image offsets were MMIO register offsets (0x30, 0x34, 0x38).
LRC state layout is different — CTX_CONTEXT_CONTROL is dword index 3 (byte 12).
Fixed to match Linux intel_lrc_reg.h dword layout:
LRC_CTX_CTRL_OFFSET = 12 (was 0x02)
LRC_HEAD_OFFSET = 20 (was 0x30)
LRC_TAIL_OFFSET = 28 (was 0x34)
LRC_RING_START_OFFSET = 36 (was 0x38)
Added LRC_RING_CTL_OFFSET = 40
- context.rs: corrected CTX_CTRL_INHIBIT_SYN_CTX_SWITCH from bit 2 to bit 3
to match Linux RING_CONTEXT_CONTROL bit layout at 0x244.
- ring.rs: added gpu_address() method to expose ring GGTT address for
LRC descriptor construction.
- mod.rs: wire execlist submission path after ring batch submit in
hw_submit_to_ring(), creating LRC descriptor and submitting via ExeclistPort.
- display.rs: program HDMI infoframes (AVI, audio, VSIF) on port enable.
On Redox there is no udev-based DRM device enumeration. KWin's DRM
backend relies on m_udev->listGPUs() which returns nothing without udev.
Add a fallback: when no KWIN_DRM_DEVICES is set and the session is
kde-wayland, inject KWIN_DRM_DEVICES=/scheme/drm/card0 so KWin knows
which device to open. This fixes the 'No suitable DRM devices' error
that prevented KWin from starting on Redox.
Added test: build_environment_sets_kwin_drm_devices_default_for_kde_wayland
to verify the fallback is applied correctly.
- gem_lmem: Replace bump allocator with best-fit free-list (BTreeMap) that
tracks individual allocations and coalesces freed blocks on both sides
- mocs: Add init_pat() - programs PAT index 0-7 with WB/WC/WT/UC for Gen9+,
and WB-only for Gen12+; called after init_mocs() in IntelDriver init
- regs_gt: Add PAT register constants (GEN8_PRIVATE_PAT_*, GEN12_PAT_INDEX,
GEN8_PPAT_* cache attributes) and TBIMR_FAST_CLIP
- PAT programming: Gen9 uses 0x40E0 base with LLC/LLCELLC attributes,
Gen12 uses 0x4800 base with simple WB/WC/WT/UC (no LLC on Xe2)
- All changes compile clean (0 errors)
Fix pre-existing compilation errors in modules that were present as
source files but not declared in mod.rs:
- audio_eld: cast u16 copy_len to usize for slice indexing
- dp_fec, dp_uhbr, edp_pll, gpu_reset, hdmi_frl, lspcon:
DriverError::Initialization now takes String, add .to_string()
- dsc: add missing import
- guc_submission: DriverError::Buffer now takes String
- vrr: cast VRR_MAX/MIN_FRAME_TIME constants to usize
- rps_rc6: change freq_table() return to &'static to avoid
borrow checker conflict with self mutation
All 12 modules now compile with zero errors.
Replace busy-wait spin_loop() in FenceTimeline::wait() with
Condvar::wait_timeout(). signal() now calls notify_all() to
wake blocked threads. This turns syncobj_wait from CPU-burning
poll to proper blocking sleep/wake.
Add two new tests:
- test_wait_wakes_on_signal: spawns a thread that signals
after 10ms, verifies the blocked wait wakes within 1s
- test_wait_timeout_expires: verifies 1ms timeout on an
unsignaled fence returns an error
VIRTGPU_GETPARAM: expand from 1 to 8 sub-parameters for Mesa
compatibility. Mesa virgl driver probes CAPSET_QUERY_FIX,
RESOURCE_BLOB, CONTEXT_INIT, SUPPORTED_CAPSET_IDS, and
EXPLICIT_DEBUG_NAME during initialization.
set_property: add doc comment explaining that virtio-gpu has
no per-object property tables — all mode/fb/active changes
flow through atomic commit, not set_property.
cursor_move: replace x.max(0) as u32 / y.max(0) as u32 with
explicit if-else for clarity. Negative coordinates now clamp
to zero at screen edges (same behavior, more readable code).
BUG 1: virgl_resource_create_blob held device lock while calling
self.gem_create() which internally tries to lock device again.
Rust std::sync::Mutex is not reentrant — guaranteed deadlock.
Fix: release device lock before calling gem_create, using a
scoped block for the has_resource_blob feature check.
BUG 2: Box::leak in atomic_commit error paths converted
dynamically-formatted strings to &'static str at the cost of
a memory leak per error. Replaced with static &str literals.
VirtioDriver:
- Override poll_hotplug() — refresh connectors and detect
display changes by comparing cached vs current topology
- Override set_property() — validate obj_id is a known CRTC
or connector; compositors need property acknowledgement
even if individual properties are no-ops for virtio-gpu
scheme.rs:
- Fix fsync() — was EOPNOTSUPP, now returns Ok(())
Virtio-gpu commands complete synchronously, so there
are no pending GPU operations to flush
Implement VIRTGPU_RESOURCE_CREATE_BLOB:
- Define VirtioGpuResourceCreateBlob wire struct (commands.rs)
- Add VIRTIO_GPU_BLOB_MEM_*/FLAG_* constants
- Negotiate VIRTIO_GPU_F_RESOURCE_BLOB feature flag
- Add virgl_resource_create_blob() to GpuDriver trait
- Implement in VirtioDriver with virtio command dispatch
- Wire ioctl handler in scheme.rs (was EOPNOTSUPP stub)
- Add find_by_handle() to ResourceManager
Implement hardware cursor:
- Add VIRTIO_GPU_CMD_UPDATE_CURSOR/MOVE_CURSOR opcodes
- Define VirtioGpuCmdUpdateCursor/MoveCursor/CursorPos structs
- Add update_cursor()/move_cursor() to VirtioGpuDevice
- Override cursor_set/cursor_move on VirtioDriver
- CRTC-to-connector lookup for scanout index mapping
Implement atomic modeset:
- Override atomic_commit on VirtioDriver with full state
validation via atomic_check(), then delegate to
set_crtc + page_flip for each active CRTC
- Support TEST_ONLY flag (returns NoChange)
Mesa recipe: add iris,crocus to gallium-drivers
Config: enable mesa = {} in redbear-full.toml
Extract protocol-agnostic FenceTimeline from Intel to shared
src/drivers/fence.rs — atomic-based fence tracking suitable
for Intel, VIRGL, and AMD drivers.
Extract protocol-agnostic SyncobjManager from Intel to shared
src/drivers/syncobj.rs — syncobj create/destroy/signal/reset/
wait/query and sync_file fd export/import.
Wire both into VirtioDriver:
- Add FenceTimeline + SyncobjManager fields
- Implement all 5 GpuDriver syncobj trait methods
(create, destroy, wait, export_fd, import_fd)
- Track fence seqnos in virgl_submit_3d (allocate
before submit, signal after completion)
Intel fence.rs and syncobj.rs converted to thin re-export
modules pointing at shared sources — no behavioral change
for Intel driver.
This gives Mesa VIRGL userspace the standard DRM syncobj
API for GPU/compositor synchronization.
mod.rs: added pub mod declarations for all previously-dead modules
PHY: mg_pll, dkl_phy, cx0_phy, snps_phy, edp_pll
Power: rps_rc6, dmc_power, psr_full, guc_submission, alpm
Display: dp_fec, dp_audio, hdmi_frl, hdmi_scrambler, dp_uhbr,
dp_phy, vrr, dsc, display_irq, bandwidth, panel_fitter
Platform: lspcon, tc_port
GT: workarounds, gpu_reset
Support: audio_eld, cdclk_tables, color_lmem, color_pipeline
handle_irq: fixed MSI-X EOI gap
InterruptHandle::eoi() now called after process_irq()
Lock held across IRQ processing to prevent early drop
Prevents MSI-X vectors from firing only once
30 modules were never compiled — now all 65 source files participate
in compilation. Legacy issues in previously-uncompiled modules remain
and will be addressed separately.
vasilito
Sisyphus