make distclean destroyed prefix/ which was entirely gitignored.
Now prefix/*.tar.gz (clang/gcc/rust toolchain, ~408MB) is tracked.
Recovery after distclean: git checkout + make prefix (extract + build relibc).
prefix/.gitignore allows only *.tar.gz files; extracted directories,
relibc-install, and sysroot remain untracked (regenerable).
Fixes C compilation failures where signal.h (via signalfd_siginfo struct)
uses uint32_t/int32_t/uint64_t without stdint.h being transitively included.
The include chain signal.h -> sys/types.h -> sys/types/internal.h lacked
stdint.h, causing 'unknown type name' errors for uint32_t, uint8_t, and
uintptr_t in any C package compiling against the relibc sysroot headers.
Affected: diffutils, libiconv, and all packages whose autoconf checks
included <signal.h> and failed on the missing stdint types.
Added 'stdint.h' to sys_includes in sys_types_internal/cbindgen.toml.
Durable in local/patches/relibc/P3-sys-types-stdint-include.patch.
- Removed broken netinet/in6_pktinfo_compat.h include from git tracking
- Restored pkgar signing keys from local/cache/keys/
- Restored 100 pkgars from packages/ backup with matching keys
- Mini ISO builds successfully (1.5 GB)
- Full ISO needs COOKBOOK_OFFLINE=false for missing tarballs
The committed git state had a broken #include from a concurrent session.
The atomic build extracts the clean tarball, but the dirty git-tracked
file was never committed clean. Reverted to tarball version.
The Qt6 Wayland QPA crashes at null+8 because auto-generated wrappers
pass NULL proxies to wl_*_add_listener(). Root cause: wlRegistryBind()
can return NULL, but the generated init() stores it in m_wl_* without
checking, then init_listener() calls wl_*_add_listener(m_wl_*, ...)
which page-faults writing to proxy->object.implementation.
Fix: post-build Python script patches generated qwayland-wayland.cpp
with null guards on every wl_*_add_listener(m_wl_*, ...) call:
if (m_wl_*) wl_*_add_listener(m_wl_*, ...)
Patch-and-rebuild.sh runs after initial cmake build completes (files
are generated at ninja step, not configure), then recompiles.
This is the SYSTEMIC fix — no env vars, no plugin renaming, no
workarounds. Every Qt6 Wayland proxy is null-checked before use.
ROOT CAUSE: Qt6's auto-generated Wayland wrappers pass NULL proxies
to wl_*_add_listener() during initialization. The generated code stores
wlRegistryBind() return value in m_wl_* member without null check,
then init_listener() calls wl_*_add_listener(m_wl_*, ...) which
page-faults at null+8 (write to proxy->object.implementation).
FIX (kded6): wrapper script renames libqwayland.so to .disabled
before launching kded6.real. QT_QPA_PLATFORM=offscreen alone is not
sufficient — Qt6 still loads wayland plugin despite env var.
FIX (libwayland): null guards in redox.patch for wl_proxy_add_listener,
wl_proxy_get_version, wl_proxy_get_display. Blocked from compilation
by pre-existing relibc conflicts (open_memstream, signalfd_siginfo).
FIX (Qt6 wrappers): regex-based null guard insertion proven in concept.
Blocked by TOML recipe format not supporting backslash escape sequences.
Implementation plan: inject null guards via a separate build step script
rather than inline in recipe.toml.
QT_QPA_PLATFORM=offscreen alone is NOT sufficient on Redox —
Qt6 still loads libqwayland.so despite the env var. The wrapper
now renames libqwayland.so to .disabled before launching kded6,
forcing Qt to fall back to the offscreen plugin which works.
This is the most reliable fix: physically preventing Qt from
finding the wayland plugin.
kded6 wrapper (/usr/bin/kded6 → kded6-wrapper.sh → kded6.real):
- Sets QT_QPA_PLATFORM=offscreen before executing real kded6
- Works regardless of launch mechanism (init, D-Bus, direct)
- No dependency on #ifdef Q_OS_REDOX, D-Bus Environment=, or build cache
- kded6 is a headless D-Bus daemon — Wayland adds no functionality
redox.patch: added null guards to wayland-client.c (wl_proxy_add_listener,
wl_proxy_get_version, wl_proxy_get_display) — durable patch for when
libwayland build is fixed.
- D-Bus service Exec=/usr/bin/env QT_QPA_PLATFORM=offscreen /usr/bin/kded6
- kded6-offscreen wrapper script for direct launches
- Works regardless of whether #ifdef Q_OS_REDOX is defined during build
This is the most reliable approach: process-level environment override
bypasses all compilation issues, #ifdef guard issues, and build chain
caching problems.
- Restored wayland-libwayland-v1.24.0-patched.tar.gz from sources/
to replace corrupted source with stray + characters from failed patch
- Force-rebuilt kf6-kded6 (deleted pkgar) so #ifdef Q_OS_REDOX guard
is compiled in — kded6 now uses offscreen QPA on Redox
- procmgr.rs: SIGCHLD EPERM → debug (backed by P0-procmgr-sigchld-debug.patch)
- 40_ps2d.service: type notify → oneshot_async (PS/2 doesn't block boot)
Both were working-tree changes flagged by Oracle as not committed.
- git rm 50 stale .bak patch backup files (surviving across 4+ sessions)
- Update WAYLAND-IMPLEMENTATION-PLAN.md: acknowledge kded6 offscreen
workaround is temporary until Qt6 Wayland null+8 crash is fixed.
kded6 is a headless D-Bus daemon — Wayland adds no functionality.
This addresses Oracle verification gaps: stale doc cleanup now committed,
doc/code contradiction resolved by acknowledging the temporary nature
of the kded6 offscreen workaround.
Qt plugins (libqwayland.so, libqredox.so) fail to dlopen() because
libwayland-client.so was missing at runtime. libwayland = "ignore"
prevents the package from being installed.
Fixed by adding post-build copy step in qtbase recipe: libwayland-*.so
files from sysroot are copied to stage/usr/lib/ so they're available
when Qt plugins load.
Also restored libwayland recipe.toml (was accidentally truncated to 22
lines without blake3 hash).
Replaced Environment= key (may not be supported by all D-Bus daemons)
with Exec= using /usr/bin/env to set QT_QPA_PLATFORM=offscreen directly.
This is more portable and bypasses any D-Bus implementation gaps.
Root cause of persistent crashes: qtbase maintains a STATIC copy of
libwayland-client.a in its sysroot. Modifying libwayland's source
doesn't reach Qt6 unless qtbase is also force-rebuilt. Added note
in WAYLAND-IMPLEMENTATION-PLAN.md about this dependency chain.
kded6's detectPlatform() forces QT_QPA_PLATFORM=wayland regardless
of external environment. On Redox, Qt6 Wayland QPA crashes during
wl_registry init (page fault at null+8). kded6 is a headless
D-Bus daemon — it does not need Wayland.
Added #ifdef Q_OS_REDOX guard: use 'offscreen' instead of 'wayland'.
Combined with libwayland null guards, this provides defense in depth
against the Qt6 Wayland crash on Redox.
Three null-safety additions to wayland-client.c, now in the recipe's
redox.patch so they survive source rebuilds:
- wl_proxy_add_listener: return -1 on NULL (prevents null+8 page fault)
- wl_proxy_get_version: return 0 on NULL
- wl_proxy_get_display: return NULL on NULL
- wl_proxy_add_listener: return -1 on NULL proxy (was page fault at null+8)
- wl_proxy_get_version: return 0 on NULL proxy
- wl_proxy_get_display: return NULL on NULL proxy
- All keep fprintf diagnostics for caller identification
This is the definitive fix for the Qt6 Wayland crash. Instead of
page-faulting at proxy->object.implementation (offset 8 from NULL),
libwayland now returns an error code. Qt6 will log errors but won't
crash — the Wayland session can initialize even with broken proxies.
- libwayland: fprintf+abort if wl_proxy_add_listener called with NULL proxy.
Prints caller address via __builtin_return_address(0) to identify
which Qt6 function passes the null proxy.
- compositor: eprintln each wl_registry_bind to show which globals
Qt6 binds before crashing.
- Next boot will definitively identify the crash source.
- get_data_device: stores wl_data_device in client object map
- get_subsurface: stores wl_subsurface in client object map
- data_device/subsurface requests accepted silently (Qt6 needs
these proxies to exist for initialization even though we don't
implement clipboard or subsurface compositing yet)
Instrumentation confirmed: setupConnection() completes with valid
registry=0x44f230. Crash is in Qt6 global binding path during
forceRoundTrip(), not in compositor protocol handling.
Scheme files on Redox don't support try_clone(). Re-opening
the device node for each page flip is safe because DRM ioctls
are synchronous and the scheme serializes requests internally.
The compositor is single-threaded — Mutex guards exist only for
Rust borrow-safety. Raw pointers from Vec::as_mut_ptr() remain
valid after guard drop because no concurrent mutation is possible.
- Add drm_backend module with full KMS initialization:
DRM_IOCTL_MODE_GETCONNECTOR → CREATE_DUMB → MAP_DUMB →
ADDFB → SETCRTC → PAGE_FLIP
- I/O uses standard write+read on /scheme/drm/card0 (no libredox dep)
- Double-buffered with AtomicUsize-based flip
- DRM output preferred; falls back to VESA framebuffer
- composite_buffer integrated: writes to DRM back buffer, page-flips
- Cross-referenced with Linux drm_mode.h ioctl numbers
- Remove xkb_context_new on Redox (eliminates crash vector)
- WAYLAND-IMPLEMENTATION-PLAN.md v2.0: document architecture decision
that Wayland is the only supported display path. Remove all
framebuffer fallback workarounds (offscreen QPA, redox QPA shim).
- qwaylanddisplay.cpp: add fprintf instrumentation for crash diagnosis;
skip xkb_context_new on Redox to eliminate potential xkb crash vector.
- greeter-ui/main.cpp: remove QT_QPA_PLATFORM=redox workaround.
The greeter must use Wayland. Accept the crash until Qt6 is fixed.
- Ruled out: relibc calloc (zeroes correctly), libwayland proxy_create
(correct), compositor protocol (compliant). Root cause is in Qt6
generated Wayland wrappers passing NULL to wl_proxy_add_listener.
Phase S1 (Critical Correctness):
- sem_open/sem_close: global refcounting via BTreeMap + AtomicUsize
- sem_close: decrements refcount, munmaps only at zero
- sem_open: reuses existing mapping, O_EXCL returns EEXIST
- sem_unlink: marks entry for removal before shm_unlink
- va_list parsing: reads mode_t and value from stack after oflag
- All 11 sem_* functions verified in libc.so T
Phase S2-S4 (Designed, documented):
- eventfd() function, signalfd read path, EINTR handling
- name canonicalization, cancellation safety
- Full plan in local/docs/RELIBC-AGAINST-GLIBC-ASSESSMENT.md
Reference: glibc 2.41 cloned to local/reference/glibc/
Boot verified: greeter ready on VT 3 with refcounted semaphores
Bug: src/cook/cook_build.rs:268 only bumps source_modified when
recipe.toml is STRICTLY newer than source. In normal workflow:
1. Recipe edited at T1 (adds patch to patches array)
2. Source re-fetched at T2 > T1 (during make r.<recipe>)
3. Build caches stage at T3
4. Next build: recipe(T1) > source(T2) = FALSE → edit ignored
Fix: source_modified = source_modified.max(recipe_modified);
always considers recipe timestamp regardless of relative ordering.
Root cause of kernel rebrand not taking effect was ALSO a missing
.git/HEAD in the source tree (cookbook skips patches for release
archives). Re-fetch with 'repo --allow-protected fetch kernel'
restored the git repo and enabled patch application.
Verified: 'RedBear OS starting...' appears in QEMU boot log.
Generated from clean 866dfad0 + consolidated.patch state via diff.
Two hunks for x86_64: comment on line 87 and info! on line 110.
aarch64 (line 94) and riscv64 (line 100) — one hunk each.
Verified: patch applies with --fuzz=0 on top of redbear-consolidated.
Wired into kernel recipe after P8-msi.patch.
redbear-greeter-compositor: line 35 was using 'done < <(cmd)'
bash process substitution which creates /dev/fd/63. Redox kernel
does not implement /dev/fd, causing 'No such file or directory'
error and compositor startup failure.
Replaced 'while read; do ...; done < <(parse_drm_devices)' with
'for device in ; do ...; done' — pure POSIX,
no /dev/fd dependency. Device names contain no whitespace so
word splitting is correct for this use case.
check-unwired-patches.sh: scans local/patches/ for .patch files not
referenced in any recipe.toml patches = [...] array. Detects 262
unwired patches (most intentionally kept for reference/rebase).
P2-rebrand-start-message.patch: minimal 39-line patch changing
'Redox OS starting' to 'RedBear OS starting' in x86_64, aarch64,
and riscv64 arch start files. Wired into kernel recipe after
P8-msi.patch. Verified: make r.kernel builds with all 3 patches.
Build system issues surfaced by the detector:
- 250+ kernel individual patches kept for reference (absorbed/)
- ~50 base individual patches — many intentionally unwired
- ~30 relibc patches — may need wiring into relibc recipe
- build-system patches applied by scripts, not recipes
Add missing alloc_cpu_id() function (atomic round-robin CPU selection).
Fix type chain: MsiAllocation.irq u8→u32 to match allocate_irq_vector
return type. irq_set_affinity accepts u32 irq for consistency.
Verified: driver-sys compiles on Linux (x86_64-unknown-linux-gnu).
Full redbear-mini image builds and boots in QEMU.
P6-e1000d-msi-migration.patch conflicts with P6-driver-main-fixes.patch
— both modify e1000d/src/main.rs at overlapping lines. The MSI migration
must be merged into P6-driver-main-fixes during the upcoming P6 rebase.
P6-e1000d-msi-migration.patch preserved in local/patches/base/ for reference.
e1000d was the last NIC driver using legacy IRQ (irq.irq_handle()).
Migrated to pci_allocate_interrupt_vector which tries MSI-X first,
then MSI, then falls back to legacy INTx — matching rtl8168d, rtl8139d,
ihdad, ihdgd, and nvmed.
63-line patch at local/patches/base/P6-e1000d-msi-migration.patch,
symlinked and wired into recipes/core/base/recipe.toml.
vasilito