Replace the blocking main render loop in kernel.pas with two new
timer-hooked units that are driven by the 1024Hz PIT interrupt:
- graphicsrefresh: calls desktop.update, uidebug.update, lvgl_handler
and video.Flush every 4 ticks (~256 FPS target).
- usbhotplug: calls usbcore.usb_check_hotplug every 1024 ticks (~1Hz).
The kernel main procedure now registers both timer hooks and enters an
idle STI/HLT loop (kernel.yield), allowing the CPU to sleep between
interrupts instead of busy-spinning.
Also removes the unused myUserLandFunction and reorders early syslog
init calls for clarity.
- Track enumerated devices per-HC via linked list (hc^.Devices)
- Add usb_clear_halt() for endpoint stall recovery (CLEAR_FEATURE)
- Add usb_remove_device() to clean up and free device records
- Add get_hc() for indexed host controller access
- Implement 'USB' terminal command showing HC info, connected devices,
speeds, class descriptors, and endpoint details
The PS/2 keyboard callback hardcoded is_down_code := true for every
scancode, so key releases were never reported. This caused keys to
repeat endlessly until the next key was pressed.
In PS/2 scancode set 1, bit 7 indicates a break (release) code.
The callback now:
- Detects break codes via (scancode AND $80)
- Masks to base code (scancode AND $7F) for key matrix lookup
- Sets is_down_code := false for break codes
- Updates modifier state (shift/ctrl/alt) before building the
key event, so modifier releases take effect immediately
Files changed:
- ps2_keyboard.pas: rewrite callback to handle make/break codes
Replace main-loop polling (poll_all, poll_keyboards, poll_mice) with
hardware interrupt handlers for all four USB host controller drivers.
Each HC driver (XHCI, EHCI, OHCI, UHCI) now:
- Registers an ISR on its PCI interrupt line via isrmanager
- Enables hardware interrupts after HC initialization
- Acknowledges interrupt status in the ISR before calling poll,
preventing interrupt storms on level-triggered PCI lines
- Uses a PollBusy re-entrancy guard to prevent ISR/inline poll races
usbcore gains a completion hook system (register_completion_hook /
fire_completion_hooks) so HID drivers (keyboard, mouse) are notified
directly from the ISR after transfer completion.
Files changed:
- usbcore.pas: completion hook infrastructure
- XHCI.pas: ISR, enable_interrupts, submit ordering fix
- EHCI.pas: ISR, enable_interrupts, status ack in ISR
- OHCI.pas: ISR, enable_interrupts, status ack in ISR
- UHCI.pas: ISR, enable_interrupts, status ack in ISR
- usb_keyboard.pas: register completion hook
- usb_mouse.pas: register completion hook
- New units: usbtypes, usbcore, usbhub, UHCI, OHCI, usb_keyboard,
usb_mouse
- usbtypes: shared USB descriptors, constants, HC abstraction,
aligned alloc, unit tests
- usb_keyboard: USB HID boot keyboard driver with HID-to-ASCII
translation, polling, unit tests
- usb_mouse: USB HID boot mouse driver with delta/scroll/button
processing, polling, unit tests
HID Layer Restructure:
- keyboard.pas: stripped to abstract API (hook + reportKeyEvent),
no PS/2 dependency
- mouse.pas: absorbed mousestate.pas, now owns
position/buttons/scroll/hooks directly
- mousestate.pas: deleted (merged into mouse.pas)
- New ps2_keyboard.pas: PS/2 scan code translation,
calls keyboard.reportKeyEvent
- New ps2_mouse.pas: PS/2 packet handling, calls
mouse.setMousePos/fireMouseEvent
LVGL Integration:
- lvgl.pas: updated to use mouse.pas directly (was mousestate),
improved keyboard input handling with proper key held/release tracking
- vterminal.pas: removed serial debug logging, dropped serial dependency
Kernel:
- kernel.pas: added USB units, USB polling in main loop, unit test
calls, ps2_keyboard/ps2_mouse init
When size=0 was passed to memset() or memcpy(), the loop bound
`0 to size-1` underflowed to $FFFFFFFF, iterating ~4 billion times
and causing an immediate page fault.
Added `if size = 0 then exit` guards to both memset() and memcpy()
in util.pas.
Also guarded all `for i:=0 to X-1` loops in strings.pas where X
could be 0 on empty string inputs:
- stringToUpper: wrap loop in `if stringSize > 0`
- stringToLower: wrap loop in `if stringSize > 0`
- stringEquals: early exit when both sizes are 0
- stringToInt: early exit when string is empty
Additionally fixed all previously identified bugs in strings.pas:
- hexStringToInt: guard against empty string underflow
- stringConcat: cache stringSize calls to avoid redundant O(n) scans
- stringTrim: clamp length to source string size
- stringSub: bounds check start/size against source length
- stringReplace: complete rewrite - was using stringEquals() for
substring matching (only matched at string end), dropped prefix
chars, had dangling assignment overwriting valid result with
uninitialized pointer, and had unused variable
- stringIndexOf: rewritten to use new stringMatchAt() helper for
correct substring matching
- stringContains: simplified to delegate to fixed stringIndexOf(),
also fixing uint32 underflow when sub was empty
- stringToInt: removed always-true `v >= 0` check on uint32
- intToString: implemented (was a stub returning ' ')
- Added stringMatchAt() helper for prefix comparison
- Added comprehensive UnitTest procedure with 77 assertions
including adversarial empty/nil/boundary inputs
refactor(stdio): replace console/terminal command system with buffer-based stdio
- Introduced stdio.pas as the unified command I/O layer.
- All 31 command procedures across 15 files now receive stdin_buf,
stdout_buf, and stderr_buf (POutBuf) instead of a single outbuf.
- vterminal.pas creates all three buffers per command invocation and
displays both stdout and stderr.
- Deleted console.pas, terminal.pas, and 9 obsolete windowing programs
(shell, splash, edit, memview, themer, netlog, vmlog, vmstate, udpcat).
- Migrated all registerCommand calls to stdio, rewrote vterminal
processCommand to use stdio.findCommand, and added proper error
routing to stderr for invalid params, bad paths, and system errors.
refactor(syslog): replace console logging with serial-backed syslog
- Introduced syslog.pas for kernel-level logging over serial/dev.
- Migrated all boot logging in kernel.pas from console.outputln/
writestringln to syslog.logln/writestringln.
- Converted 19 fault handlers, hashmap, lists, scheduler, tracer, net,
udp, and all driver init paths to use syslog.
- Rewrote BSOD in util.pas (50+ calls) from console to syslog.
- Removed CONSOLE_SLOW_REDRAW, TRGB565, TRGB565Pair, and HWND
from system.pas.
fix(keyboard): add ring buffer and proper press/release cycling for LVGL input
- Replaced the single last_key/key_pressed slot in lvgl.pas with a
16-entry ring buffer (kb_buf, kb_head, kb_tail) to prevent ISR
overwrites when typing fast.
- Added a kb_pending_release flag so each keystroke gets a
full PRESSED->RELEASED cycle before the next key is consumed —
LVGL requires this pairing and was silently dropping intermediate
keys when receiving consecutive PRESSED events without
intervening releases.
feat(vterminal): use Hack monospace font for terminal output
- Added Hack Regular 14px (hack_14.c) as a custom LVGL font.
- Removed the .static_bitmap field from the generated font file
(not present in LVGL 9.2 lv_font_t struct).
- Declared the hack_14 external symbol in lvgl.pas and switched
vterminal's text label from lv_font_montserrat_14 to hack_14
for proper monospace terminal rendering.
XGETBV/XSETBV require CR4 bit 18 (OSXSAVE) to be set before execution,
otherwise the CPU raises #UD (invalid opcode). Hyper-V enforces this
strictly; VirtualBox was silently tolerating it.
- Check both AVX and XSAVE CPUID flags before attempting AVX enable
- Set CR4.OSXSAVE (bit 18) before issuing XGETBV/XSETBV
- If XSAVE is not exposed by the hypervisor, AVX enable is skipped
gracefully instead of faulting
- Add mouse_wait_long() with 100,000 iteration timeout for init-time
i8042 controller commands (mouse_write/mouse_read/load). The original
mouse_wait() stays at 100 iterations for ISR-safe use in main().
- In load(): use mouse_wait_long() for all direct port I/O ($64/$60)
and clear CCB bit 5 (AND NOT $20) to ensure the mouse clock is not
disabled — required for Hyper-V which enforces this strictly.
- Replace Console.getConsoleProperties^.Width/Height with
video.frontBufferWidth/Height for mouse bounds clamping. The old
console properties are zero when console.init() is not called,
which locked the cursor to (0,0). Add 'video' to uses clause.
The LVGL flush callback assumed 32bpp (4-byte stride, direct pixel
copy), causing corrupted output on hypervisors that fall back to
16bpp modes (e.g. HyperV at 1600x1200x16).
- lvgl.pas: Make lvgl_flush_cb BPP-aware; add ARGB8888->RGB565
conversion path for 16bpp, keep direct copy for 32bpp. Replace
static lv_buf1 array (hardcoded 1600px wide) with kalloc'd buffer
sized to actual screen width.
- doublebuffer.pas: Fix allocateBackBuffer size from (W*H*BPP) to
(W*H*BPP) div 8 to get correct byte count.
- vesa.pas: Account for BPP in allocateVESAFrameBuffer page range
calculation to avoid under-mapping at higher BPP/resolutions.
- Removed the initialization of console in kernel.pas as this will
cause issues on some systems that don't expect to need to allocate the
textmode buffer.
t3hn3rd
Aaron
admin