| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The real-hardware crash (2 lines of white + freeze, no CPU exception
even with the 0.5.2 fault dumps) pointed away from a fault and at a
stall: the scroll's full-frame push used one 1.92 MB rep-movsb to the
framebuffer. On QEMU the framebuffer is RAM, so it flies; on a real
GPU the framebuffer is uncached MMIO and the CPU's fast-string burst
path can stall forever — the shell wedges mid-instruction, the cursor
task's own writes wedge too, no exception fires (MMIO writes don't
fault), and the screen shows the torn partial push (the white bands).
The boot-time clear() always worked on the same hardware and uses
per-word volatile stores, so the scroll push now uses the identical
pattern: shadow reads (cached RAM, cheap) + individually-posted
display stores. Interruptible between stores, never a bulk burst,
and functionally identical on QEMU (all 5 suites re-pass).
Also converts the no-shadow fallback shift to the same per-word
pattern. CONTEXT.md updated (scroll-cost note + versioning).
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The real-hardware scroll crash (white bands + instant freeze, QEMU-clean)
was invisible because the only trace was a single serial line on machines
with no serial console. This makes every crash self-identifying and
hardens the real-hardware failure surfaces:
- Every CPU exception vector 0-31 now has an IDT entry (macro-generated
stubs; 8/14 keep their IST stacks). A fault in the shell prints a
white-on-red band to the framebuffer via gop::dump_lines (raw fb
writes bypassing the vt, fully clamped) plus a parseable serial line:
'!!! <kind> vec=N err=… rip=… cr2=… rsp=… rflags=… task=N "name"',
then halts — instead of converging to a cryptic double fault.
Background-task faults abort the task as before.
- Boot log now reports the authoritative post-SetMode geometry:
'gop: fb base=… size=… pitch=… WxH bpp=… bgr=…', so the next real
machine's report identifies whether pitch*height exceeds the mapping.
- Framebuffer writes (clear, push_rect, shift_up_n bulk push, cursor
overlay) are clamped to the firmware's framebuffer_size: a pitch-
padded mode degrades to a truncated view instead of faulting past
the mapping.
- The GOP framebuffer region is excluded from the page allocator (and
subtracted from the heap's conventional-region pick) even when the
firmware lists it as ConventionalMemory — kernel allocations can
never overwrite the display again. The GOP protocol is located and
peeked before the memory map is captured so the exclusion is known
at heap-init time.
- Shell gains a hidden 'fault' command (deliberate unmapped write) and
'make test-fault' verifies the dump: vec=14, cr2=0x400000000000,
task=0, halt — no triple fault. All other suites re-pass.
- CONTEXT.md: fault-dump format documented for real-hardware reports.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The kernel used to inherit whatever GOP mode the firmware picked
(QEMU/OVMF gave 1280x800). Now every surface derives from a fixed
800x600 canvas:
- uefi: typed EfiGopQueryMode/SetMode/GetMode fn pointers in
GraphicsOutputProtocol (were raw *mut c_void, never called).
- gop::select_gop_mode: scan the mode list for an exact 800x600
(32-bit BGR/RGB only, the driver assumes 4 Bpp), SetMode it; fall
back to the smallest mode that fits the canvas, then to the
firmware's choice. SetMode may move both the mode info pointer and
the framebuffer base, so both are re-read after the switch. Boot
log: 'gop: mode 3 selected — 800x600 (was 1280x800)'.
- tests: relative assertions are resolution-agnostic; test-gfx and
test-mouse had a boot-log race (shell line prints before the cursor
task line) — now wait for the cursor line instead of grepping once.
All four suites pass at 800x600. 0 build warnings.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The first visible GUI surface: real rect primitives, a single
shadow→display push path, and an on-screen mouse cursor that can
never race the text surface.
gop:
- fill_rect + push_rect(x,y,w,h): every raster op rasterises into the
RAM shadow, then push_rect copies the rect to the display (write-only,
still never read) and composites the cursor on top. blit_glyph and
scroll shifts now route through it.
- cursor overlay: 12x16 pointer (const-fn built from column lists) drawn
white with a black offset outline; position in one AtomicU64 (x:high
32, y:low 32) — unpack masks the low word (a bare 'as usize' cast is
a no-op on 64-bit). The cursor never enters the shadow, so erasing is
just a shadow push of the old rect.
- cursor_redraw after scroll shifts and once at spawn (the vt diff never
re-blits idle cells).
cursor task (task 2, spawned at boot when a mouse is present):
- owns the PS/2 ring; polls every 10 ms, moves the cursor (clamped),
tracks buttons, and forwards every event to a task's queue when
'mouse <id>' armed it (the window-manager input path).
- /dev/mouse is now a position device (x/y/buttons snapshot); 'mouse'
watches the cursor live (prints on change); 'cursor' prints one-shot
state; 'gfx' draws a status-bar demo; 'gfx check' is an in-kernel
pixel self-test (shadow readback, display readback, cursor overlay
purity) — PASS/FAIL per stage.
Tests (headless QEMU + monitor injection):
- make test-gfx (new): all four compositor checks PASS.
- make test-mouse (rewritten): cursor tracks mouse_move deltas, buttons
ride along, /dev/mouse snapshot, live watch, real-event queue routing.
- make test-msg: task numbering shifted (cursor is task 2; msgrecv is
task 3) — updated, still PASS.
- make test-disk regression re-passes. 0 build warnings.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Warning cleanup (cargo check + clippy now 0 across the workspace):
- Migrate all 21 static_mut_refs sites to &raw mut accessors (fat, gop,
ps2, vt, ahci, interrupts, memory) so edition 2024 won't deny them
- panic = "abort" profiles: bare cargo check/build now works without
nightly/-Zbuild-std
- Wire TICKS in the timer ISR (it was never incremented, so the vt scroll
batching trigger was dead code)
- Clippy: efi_main becomes unsafe extern "efiapi" with a # Safety doc,
KeyBuf::new uses inline-const array repeat (no interior-mutable const),
editor main loop is a while-let, PIT control byte regrouped
- [workspace.lints.clippy] allows result_unit_err + needless_range_loop
(deliberate kernel style); large_enum_variant allowed on LineInput
- Drop now-stale static_mut_refs allow from the rsdos bin
FAT32 correctness fixes (both pre-existing, exposed by the SFN fix):
- Unique 8.3 short names: to_8_3() never disambiguated collisions, so
files sharing a long-name prefix got identical SFNs and fsck reported
duplicate directory entries. Add dir_short_names() + unique_short_name()
(~N tails, Windows-style) and force LFN when the SFN changes.
- Multi-cluster directories: scan_dir/dir_short_names stopped at the first
0x00 slot, hiding every entry written into a chained cluster (writes
failed with 'cannot write', duplicate SFNs accumulated); add_dir_entry
now packs the cluster (0xE5 deleted-markers) before extending so other
FAT readers (mtools, Windows) see the continuation too.
Other: delete dead fields (Disk.fis, Volume.fat_count), drop timer_irq/
keyboard_irq/rendered, refresh_and_exit is a documented unsafe fn, CONTEXT
.md corrected (stable toolchain, IRQ input, 4-crate layout), .gitignore
excludes .pi-subagents artifacts.
Validated: cargo check/clippy/fmt clean, make test-disk PASS on fresh and
reused images (incl. two directory-chain extensions and ~10 SFN tails),
editor round-trip smoke test PASS.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
- ahci.rs: polling AHCI on PCI class 1/6/1, bus mastering + ABAR map,
port 0 up, IDENTIFY + READ/WRITE DMA EXT through one 4KiB DMA buffer
- fat.rs: BPB parse, whole-FAT RAM cache with write-back to both copies,
LFN read/write, dirs, multi-cluster files, FSInfo free-count sync,
space-padded dot entries ('..' = cluster 0 under root), fsck-clean
- fs.rs: /disk mount node + 24-slot ephemeral node pool; ls/cat/hexdump/
echo/mkdir/touch/rm/edit all work on disk paths unchanged
- QEMU: explicit bus assignment (disk=ide.0, ESP=ide.1); the implicit
fat:rw:esp drive was stealing AHCI port 0
- Debugged against QEMU source: header DW0 is CFL in bits 4:0 / PRDTL in
bits 31:16 (Linux libahci layout), FIS count at bytes 12-13, device at
byte 7; QEMU rejects H2D FIS flags bits 6:4, so direction is the
header W bit only
- Verified: full shell round-trip on /disk, reboot persistence (poweroff
then mdir/mtype and a fresh boot), fsck.fat clean
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Streaming output scrolled one framebuffer push per line (each a
~3-8 MB write-only copy). The vt now scrolls the logical grid
instantly in RAM and defers the display shift: N pending scrolls are
pushed together as one shift_up_n — one push regardless of batch size.
- vt: PENDING_SCROLLS counter; scroll_up shifts CELLS only; flush
renders nothing while pending (the diff would otherwise re-rasterise
the whole screen per line); sync_scroll() does one gop::shift_up_n +
RENDERED mirror + normal diff flush (which re-blits just the newly
exposed rows and the cursor).
- Sync triggers: batch of 16 rows, ~40 ms of stream pause (100 Hz
TICKS), any CSI sequence (full-screen apps like the editor render
immediately), clear_all/set_cursor_visible, interactive input
(tty read_line/read_byte), and the shell before each prompt.
- gop: shift_up → shift_up_n(n) (one shadow memcpy + one push).
Verified in QEMU: mmap/hexdump streaming renders in coherent batches
with the prompt syncing at the end; clear wipes clean; the editor
scrolls and redraws immediately (45-line file, gutter + status bar
intact, cursor correct); save/exit/cat round-trip unchanged. Release
and dev builds pass.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Real-hardware GOP framebuffers are typically uncacheable or
write-combining (firmware MTRR/PAT), so the scroll shift's full-screen
memcpy *read* ~4-8 MB through uncached memory — a bus round-trip per
cache line. That made scroll printing crawl on real hardware while
plain printing (a few glyph writes) stayed instant.
- gop keeps a RAM shadow copy of the whole framebuffer (same pitch
layout), allocated at init.
- Every raster write (glyph blits, clear) goes to both the display and
the shadow.
- shift_up becomes: cached shadow memcpy up one row + clear exposed
bottom row in the shadow + one write-only push shadow→display via
copy_nonoverlapping (64-byte bursts that coalesce on WC memory) —
the display is never read.
- Fallback to the old direct shift if the shadow allocation is null.
Verified in QEMU: mmap output scrolls ~100 lines with coherent rows
and a clean bottom row, clear wipes the screen, cursor tracks the
prompt, editor paths untouched. Release + dev builds pass.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The editor's gutter '│' (U+2502) appeared as '???' — write_bytes fed
each UTF-8 byte to the state machine, which mapped every byte >= 0x80
to '?', and the Latin-1 font has no box-drawing glyphs anyway.
- vt.write_bytes now decodes UTF-8 sequences; printable non-ASCII
dispatches to the glyph path instead of the byte parser.
- gop: unicode_glyph_bits() procedurally rasterises box-drawing
(U+2500-257F light/heavy/double), arrows (U+2190-2193) and the em
dash (U+2014) — pure geometry, no font data needed; blit_glyph falls
back to it for ch >= 256.
Verified via framebuffer dumps: editor gutter │ renders as a vertical
line, help screen ═ separators and ←→↑↓ all render as lines (no '?').
Shell banner em dash renders too.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Kernel now has the classic UNIX console split, so full-screen ANSI apps
can run in ring 0:
- vt.rs: terminal emulator owning the text screen — cell grid (char +
fg/bg), logical cursor (inverse-video), scrolling, and a small ANSI
CSI parser (cursor position H/f, erase J/K, moves A-D, SGR 16-colour
with correct ANSI→VGA mapping, cursor show/hide ?25). Blits only
changed cells; scroll shifts cell grids + pixels and re-rasterises
only the exposed bottom row. gop_println!/clear_screen now render
through the vt; gop.rs is the raw rasteriser (colored glyph blits,
pixel shift).
- tty.rs: line discipline — canonical read_line (echo, line buffer,
^C cancel, ^D EOF) moved out of the shell, and raw read_byte that
translates KeyEvents to the terminal byte protocol (arrows → ESC[A-D,
Backspace → 0x7f, ^X → 0x18, Ctrl+arrows → ESC[1;5A-D, Home/End/PgUp/
PgDn/Delete). interrupt_pressed() stays for streaming commands.
- ps2: generic Ctrl(char) + CtrlArrow events (was: only CtrlC/CtrlD).
- rsdosedit: no_std port of the RNG rsedit editor as the first real
app — undo, syntax highlighting (verified green/gray/yellow on the
framebuffer), search, word nav, smart indent, ^S save / ^X exit /
^X again to discard. termios → tty raw, stdout → vt, std::fs → VFS
(ramfs), SIGWINCH → fixed geometry, force-exit window → key counter.
- shell: 'edit <path>' command; read_line/strbuf moved to rsnix.
- CONTEXT.md: layout + milestones + known-issue perf notes.
Verified in QEMU: full edit → type → arrow-move → insert → save → exit
round trip with byte-exact file content; ANSI colors via screendump.
|
|
|
Architecture:
rs-dos/ (workspace root, was raisernics/)
├── rsnix/ — kernel lib crate (uefi, serial, gop, ps2, interrupts, memory)
├── shell/ — shell lib crate (prompt, commands, line input)
├── rsdos/ — binary crate (efi_main, init, panic handler)
└── rsboot/ — future git submodule
No functional changes — boots to shell prompt identically.
Version auto-pulled from Cargo.toml via env!("CARGO_PKG_VERSION").
|
