BITS 16

global pread_sector

extern boot2_main

%define BIOS_MEMDETECT_MAGIC 0x534d4150
%define MMAP_BASE 0x1000

start:
  mov si, hallo
  call print

  push dword MMAP_BASE
  call near detect_memory
  add esp, 4

  ; hide cursor
  mov ah, 0x01
  mov cx, 0x2607
  int 0x10

  jmp enter_protected

hallo: db 'xos boot loader stage 1 ...', 0xa, 0xd, 0

print:
  mov ah, 0x0e
.repeat:
  lodsb
  cmp al, 0
  je .done
  int 0x10
  jmp .repeat
.done:
  ret

enter_protected:
  cli
  lgdt [gdt_desc]

  mov eax, cr0
  or al, 1
  mov cr0, eax

  jmp 0x08:protected_main

detect_memory:
  mov di, [esp + 2] ; mmap base
  add di, 4 ; first dword reserved for count
  xor ebp, ebp ; valid entries counter

  xor ebx, ebx

  call .e820
  jc .failed

  ; first call checks:
  cmp eax, BIOS_MEMDETECT_MAGIC
  jne .failed
  test ebx, ebx
  je .failed

.item:
  jcxz .next
  cmp cl, 0x18 ; extended record?
  jb .normal
.extended:
  test [di + 20], byte 1 ; ignored
  je .next
.normal:
  mov ecx, [di + 8]
  or ecx, [di + 12] ; zero length?
  jz .next
  inc bp
  add di, 20

.next:
  test ebx, ebx
  je .finish
  call .e820
  jc .finish
  jmp .item

.failed:
  mov ebp, 0
.finish:
  mov di, [esp+2]
  mov [di], ebp
  ret

.e820
  mov eax, 0xe820
  mov edx, BIOS_MEMDETECT_MAGIC
  mov [di + 20], dword 1
  mov ecx, 24
  int 0x15
  ret

bits 32

protected_main:
  mov eax, 0x10
  mov ds, ax
  mov es, ax
  mov ss, ax
  mov eax, 0x8000
  mov esp, eax

  jmp boot2_main

; pread_sector(unit32_t lba, uint8_t *dest)
%define DRIVE_PORT_BASE 0x01f0
%define DRIVE_PORT(n) (DRIVE_PORT_BASE + n)

%macro PORT_OUTB 2
  mov dx, %1
  mov al, %2
  out dx, al
%endmacro

pread_sector:
  push ebp
  mov ebp, esp
  push edi

  mov ecx, 8
  mov dx, DRIVE_PORT(7)
  in al, dx
  in al, dx
  in al, dx
  in al, dx

  mov eax, [ebp + 8]
  shr eax, 24
  and eax, 0x0f
  or al, 0xe0
  PORT_OUTB DRIVE_PORT(6), al ; drive select MASTER | 4bits of LBA high

  mov ecx, 4
  mov dx, DRIVE_PORT(7)
.delay:
  in al, dx
  loop .delay

  ;PORT_OUTB DRIVE_PORT(1), 0x00 ; legacy
  PORT_OUTB DRIVE_PORT(2), 0x01 ; sector count

  mov eax, [ebp + 8]
  and eax, 0xff
  PORT_OUTB DRIVE_PORT(3), al ; LBA low
  mov eax, [ebp + 8]
  shr eax, 8
  and eax, 0xff
  PORT_OUTB DRIVE_PORT(4), al ; LBA mid
  mov eax, [ebp + 8]
  shr eax, 16
  and eax, 0xff
  PORT_OUTB DRIVE_PORT(5), al ; LBA high

  PORT_OUTB DRIVE_PORT(7), 0x20 ; PIO READ

  mov dx, DRIVE_PORT(7)
.wait_rdy:
  in al, dx
  and al, 0x80
  jnz .wait_rdy

  mov dx, DRIVE_PORT(0)
  mov edi, [ebp + 12]
  mov ecx, 0x80 ; 128 x 4 byte = 512 byte block
.read_loop:
  in eax, dx
  stosd
  loop .read_loop

  pop edi
  pop ebp
  ret

bits 16

gdt_desc:
  dw 0x1f
  dd gdt

gdt:

gdt_entry0: ; invalid
  .limit_low:   dw 0x42
  .base_low:    dw 0
  .base_middle: db 0
  .access:      db 0 ;b10011010
  .granularity: db 0 ;b11000000
  .base_high:   db 0

gdt_entry1: ; flat code
  .limit_low:   dw 0xffff
  .base_low:    dw 0
  .base_middle: db 0
  .access:      db 0b10011010
  .granularity: db 0b11001111
  .base_high:   db 0

gdt_entry2: ; flat data
  .limit_low:   dw 0xffff
  .base_low:    dw 0x0000
  .base_middle: db 0
  .access:      db 0b10010010
  .granularity: db 0b11001111
  .base_high:   db 0