c55e65de5e
Records the measured effect of the region-based allocator (2.2x on a cold ls, 6 s off the boot), the two pre-existing UEFI blockers found while verifying it, and the three confounds that each produced a confident wrong answer about guest speed in one session. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
105 lines
5.3 KiB
Markdown
105 lines
5.3 KiB
Markdown
---
|
|
name: schos-uefi-loader
|
|
description: "boot/uefi works under OVMF; the SCH64 entry, the GDT trap, and the one open bug (e820 fragmentation)"
|
|
metadata:
|
|
node_type: memory
|
|
originSessionId: d4821aa9-2e49-40da-91e7-90ef9ed9997e
|
|
modified: 2026-08-09T19:05:15.325Z
|
|
---
|
|
|
|
**`boot/uefi/` boots.** `make uefi`, `make esp`, `make qemu-uefi`. Verified
|
|
under OVMF 2026-08-09: kernel loaded, PCI/VGA/e1000/RTC up, reaches
|
|
`cpu0: starting`.
|
|
|
|
Written from the UEFI specification, not gnu-efi or EDK II — a large dependency
|
|
for ~200 lines of use, and a licensing question ([[schunix-licensing]]) the
|
|
project does not need. clang emits PE/COFF directly
|
|
(`--target=x86_64-unknown-windows`, `lld-link -subsystem:efi_application`), so
|
|
**no new toolchain**. `-mno-red-zone` is mandatory: firmware can interrupt at
|
|
any point and would clobber the red zone.
|
|
|
|
## The second kernel entry
|
|
|
|
The multiboot entry is **32-bit** code that climbs into long mode itself. UEFI
|
|
hands over long mode already, and dropping back down to climb out again is a
|
|
mode switch whose failure mode is a triple fault with no console. So entry.S
|
|
publishes a 64-bit entry the same way it publishes the multiboot header: magic
|
|
`SCH64BOO`, 8-byte aligned in the first 16 KB, then a 64-bit physical address.
|
|
`include/schboot.h` owns the constants so loader and kernel cannot disagree —
|
|
the first attempt defined them twice and searched for `SCH6` while the
|
|
assembler had written `6HCS`.
|
|
|
|
**Put that header *after* the multiboot header, never before.** `header_addr`
|
|
claims the multiboot header sits at `mboot_load_addr`, and the BIOS loader
|
|
computes the kernel's file offset from `(header_addr - load_addr)`. Bytes
|
|
inserted ahead of it load the kernel shifted, which fails as a triple fault.
|
|
|
|
Entry expects: long mode, interrupts off; CR3 mapping the low 4 GB **and** the
|
|
same 4 GB at `0xffff800000000000` (entry64low's first instruction is a movabs
|
|
to a high address — identity alone triple faults); RBX = framebuffer physical
|
|
address, never 1 (that marks an AP).
|
|
|
|
**The kernel loads its own GDT there (`entry64uefi`), not the loader.** A loader
|
|
arriving in long mode is still on the firmware's descriptors; the first thing
|
|
needing the kernel's own selectors faulted with `unexpected trap 13 ... cs=38`,
|
|
a long way from the cause.
|
|
|
|
## The memory map — fixed 2026-08-09
|
|
|
|
`e820_init` used to take the usable run **containing the kernel**, which under
|
|
OVMF ends at 8 MB because the firmware puts an ACPI NVS region at 0x800000.
|
|
That failed the `PHYSTOP_MIN` check, so the kernel fell back to asserting
|
|
224 MB it had not measured and then freed straight through the reserved hole.
|
|
|
|
The kernel now keeps the usable regions as a **list** (`mem_usable[]` /
|
|
`nmem_usable` in memlayout.h) as well as a top address, and `kinit1` frees each
|
|
one clipped to the window `main()` passes in. `PHYSTOP` still means exactly what
|
|
it did, so no call site changed. `phystop` comes from the **largest** region,
|
|
not the highest, because the DMA arena is carved off the top of it and must be
|
|
contiguous; a smaller region above it is deliberately left unfreed rather than
|
|
raising the bound past a hole.
|
|
|
|
Result: 1018 MB in 7 regions under OVMF, 1023 MB in 2 under BIOS.
|
|
|
|
**It is also a large speedup, which was not the goal and is worth knowing.**
|
|
Measured from inside the guest (`/proc/uptime`, fresh image each boot,
|
|
alternating rounds): boot to prompt **14.8 s vs 21.2 s**, and a cold
|
|
`ls /bin | wc -l` **23.6 s vs 51.2 s** — 2.2x. Both rounds agreed within
|
|
0.3 s. So if the allocator ever looks slow again, this is the baseline to
|
|
compare against.
|
|
|
|
## What still stops a full UEFI boot
|
|
|
|
Two separate, pre-existing problems — neither is the memory map, and both
|
|
predate the e820 work (the unfixed kernel triple faults in the same place):
|
|
|
|
1. **virtio-gpu's BAR lands above 4 GB.** OVMF assigns the 64-bit BAR high;
|
|
the first MMIO write faults with `CR2` around `0xffff80c000000014`, which is
|
|
`P2V` of physical **768 GB**, far outside the 4 GB direct map `kvmalloc`
|
|
builds from four 1 GB pages. Then a double fault, then a triple fault and a
|
|
reboot loop. SeaBIOS assigns it low, so the BIOS path never sees it.
|
|
**Drop `-device virtio-gpu-pci` to boot UEFI past this.**
|
|
2. **Boot stalls after `cpu0: starting`**, before `schfs: mounted`. Under UEFI
|
|
the ESP is hd0 and schfs.img is hd1; the root never mounts. Not diagnosed.
|
|
|
|
`cpu0: starting` remains the UEFI high-water mark, now reached with a truthful
|
|
memory map rather than a fabricated one.
|
|
|
|
## Testing
|
|
|
|
QEMU ships the firmware: `share/edk2-x86_64-code.fd` plus
|
|
`edk2-i386-vars.fd` as the varstore, copied because it must be **writable**
|
|
(a copy on Windows inherits the read-only attribute — "Block node is
|
|
read-only").
|
|
|
|
- **i440fx, not q35.** The kernel's IDE driver speaks the legacy ports at
|
|
0x1f0; q35 has only AHCI, so the root volume would not be there at all.
|
|
- `-drive file=fat:rw:build/esp` serves a directory as FAT, so no mtools is
|
|
needed. **Do not add `format=raw`** — it conflicts with the `fat:` driver and
|
|
fails with "Block node is read-only".
|
|
- OVMF drops to its shell rather than auto-booting, so the ESP carries
|
|
`startup.nsh`. It is a **file** (`boot/uefi/startup.nsh`), not generated by
|
|
`printf`: printf reads `\E` as ESC and silently corrupts the path.
|
|
|
|
Still to do: a real FAT ESP image for `make iso`, so the ISO carries an El
|
|
Torito EFI entry beside the BIOS one.
|