DByte Native Kernel IRQ I/O

v11.7.0 extends the single-core CPL0 scheduler with a real interrupt-driven raw keyboard path. It retains the canonical PUSHAD plus EIP/CS/EFLAGS resume image and adds SEMAPHORE_WAIT as request kind 3 on the existing int 0x81 blocking boundary. No second context ABI or textual assembly is introduced.

The counting Semaphore has a bounded eight-task FIFO waiter queue and a checked maximum count. A wait consumes an existing token atomically or blocks the current task. A post wakes exactly the oldest waiter, otherwise stores one token; posting at the maximum fails without mutation. Posting preserves the incoming interrupt flag and never schedules, acknowledges hardware, or sends EOI.

The keyboard driver owns a static 64-byte FIFO ring and semaphore. It configures the i8042 with bounded status polling, installs a DPL0 interrupt gate at vector 0x21, and unmasks only PIC IRQ0 and IRQ1. The IRQ1 handler reads one raw byte from port 0x60, pushes it without overwrite, posts the semaphore, and sends exactly one master PIC EOI. It performs no allocation, mapping, logging, blocking, or scheduling.

The acceptance proof injects 1E 30 2E 20 through i8042 command 0xD2 after the reader is blocked. QEMU must deliver four real IRQ1 interrupts, the reader must resume with the exact FIFO bytes, and another runnable task must advance while input is unavailable. Separate probes cover sticky tokens, FIFO waiter order, maximum-count rejection, byte 65 drop, complete reclamation, and exact slot/frame reuse.

Run deterministic object/link checks and both 32/64 MiB QEMU boots with:

powershell -ExecutionPolicy Bypass -File scripts/verify_native_irq_io_core.ps1

Ignored ELF, BIN, disassembly, hashes, serial, debugcon, and QEMU trace evidence is written under tmp/native_v11.7.0/.

SOURCE: DBYTE_NATIVE_KERNEL_IRQ_IO.md