BITS 16
    org 0x7e00

section .code
BITS 16
    cli
    jmp gdt32Setup

align 8
gdt32:
.Null:
    dq 0000000000000000000000000000000000000000000000000000000000000000b
.Code:
;   dq    6         5         4         3         2         1         0
;   dq 3210987654321098765432109876543210987654321098765432109876543210
;   dq baseaddrGS00segl1plDXCRAbaseaddressbaseaddressbaSegmentlimitsegm
    dq 0000000011000000100110100000000000000000000000000000000011111111b
.Data:
;   dq    6         5         4         3         2         1         0
;   dq 3210987654321098765432109876543210987654321098765432109876543210
;   dq baseaddrGS00segl1plDXEWAbaseaddressbaseaddressbaSegmentlimitsegm
    dq 0000000011000000100100100000000000000000000000000000000011111111b
gdt32Pointer:
    dw  $ - gdt32 - 1
    dd  gdt32

gdt32Setup:                 ; PDP -> PD -> PT -> PE
                        ; 500   1500  2500  3500
                        ;1000   2000  3000  4000
    mov ax, 0x100
    mov ds, ax
    mov ax, 0x2000 | 0x1
    mov [ds:0], ax              ; point to PD at 0x2000

    mov ax, 0x200               ; PD is at 0x2000
    mov ds, ax
    mov ax, 0x3000 | 0x3
    mov [ds:0], ax              ; point to PT at 0x3000
    mov ebx, eax

    mov ax, 0x300               ; PT is at 0x3000
    mov ds, ax
    mov ax, 0x400               ; page entries start at 0x5000
    mov es, ax              ; offset now 0x3500

    xor esi, esi                ; starting address for identity mapping
    xor edi, edi                ; offset of page table into 0x3000
    xor ebx, ebx                ; offset of page directory into 0x2000
    mov ecx, 511                ; number of pages to assign (two megabytes for now)
    mov ebp, 0x4000             ; page table starts at 0x4500
.nextPageTable32:
    mov eax, ebp
    or al, 111b             ; set present, r/w, user flags
    mov [ds:ebx], eax           ; mov our pointer to a page entry into the page table
    add ebx, 8
    add ebp, 8
    mov eax, esi
    or eax, 111b
.nextPage32:
    mov [es:edi], eax
    add eax, 0x1000
    add esi, 0x1000
    add edi, 8
    sub ecx, 1
    jz .doneMapping32
    test edi, 0xfffff000
    je .nextPageTable32
    jmp .nextPage32

.doneMapping32:
    mov ax, 0x0
    mov ds, ax
    lgdt [gdt32Pointer]

    mov eax, 0x1000
    mov cr3, eax

    mov eax, cr0
    or eax, 0x00000011      ; set CR0.PE
    mov cr0, eax            ; enter protected mode

    jmp 1000b:now_in_prot       ; SI=1, TI=0, RPL=0, i.e. 000000000001000b

BITS 32
now_in_prot:
    jmp start32         ; should be 0x7eb2

ALIGN 8
gdt64:
.Null:
    dq 0000000000000000000000000000000000000000000000000000000000000000b
.Code:
;   dq    6         5         4         3         2         1         0
;   dq 3210987654321098765432109876543210987654321098765432109876543210
;   dq baseaddrGDLAsegl1pl11CRAbaseaddressbaseaddressbaSegmentlimitsegm
;   dq _________0L_____Ppv11C__________________________________________
    dq 0000000000100000100110000000000000000000000000000000000000000000b
.Data:
;   dq    6         5         4         3         2         1         0
;   dq 3210987654321098765432109876543210987654321098765432109876543210
;   dq baseaddrGS00segl1plDXEWAbaseaddressbaseaddressbaSegmentlimitsegm
;   dq ________________Ppv___W_________________________________________
    dq 0000000000000000100000100000000000000000000000000000000000000000b

gdt64Pointer:
    dw  $ - gdt64 - 1
    dq  gdt64



start32:                ; setup a 64-bit PML4T etc.
                    ; PML4T -> PDPT -> PDT -> PT
                    ; 0x8000  0x9000 0xa000 0xb000
    mov ax, 0x10            ; select data segment
    mov ds, ax
    mov es, ax
    mov fs, ax
    mov gs, ax
    mov ss, ax

    mov edx, 0x8000
    mov ebx, 0x9000

    mov ecx, ebx            ; ecx = phys address of PDPT
    or bl, 111b
    mov [edx], ebx          ; move pointer to PDPT into PML4 i.e. 0x9007 intno 0x8000

    mov ebx, 0xa000         ;
    mov ebp, ebx            ; ebp = phys address of PDT
    or bl, 111b
    mov [ecx], ebx          ; move pointer to PDT into PDPT i.e. 0xa007 into 0x9000

    mov ecx, 0x511          ; allocate 2 megabytes of memory
    xor edx, edx            ; page number
    mov esi, 0xa000         ; start address of page directory table
    mov edi, 0xb000         ; start address of page table

.nextPageTable64:
    mov ebx, edi            ; move address of page table into ebx
    or bl, 111b         ; set user, write, present
    mov [esi], ebx          ; move the PDTE into the PDT
    add esi, 8          ; next PDTE
    mov eax, edx            ; move first page number into eax
    or eax, 111b            ; set user, write, present
.nextPage64:
    mov [edi], eax          ; move page into page table
    add eax, 0x1000         ; next page
    add edx, 0x1000         ; next address to map
    add edi, 8          ; next address of page table entry
    sub ecx, 1
    je .doneMapping64
    test edi, 0xfffff000
    je .nextPageTable64
    jmp .nextPage64

.doneMapping64:
    mov eax, cr4            ; enable PAE paging
    bts eax, 5
    mov cr4, eax

    xchg bx, bx

    mov eax, 0x8000         ; move pointer to PML4T
    mov cr3, eax            ; store it

    mov ecx, 0xc0000080     ; EFER MSR number
    rdmsr               ; read efer
    bts eax, 8          ; set LME=1
    wrmsr               ; write efer

    mov eax, cr0
    bts eax, 31
    mov cr0, eax

    jmp 0x08:start64

BITS 64
start64:                ; should be 0x7f82
    lgdt [gdt64Pointer]
    mov rax, 0x8000
    mov cr3, rax

halt:
    hlt
    jmp halt

<bochs:2> info gdt
Global Descriptor Table (base=0x0000000000007ec8, limit=23):
GDT[0x00]=??? descriptor hi=0x00000000, lo=0x00000000
GDT[0x01]=Code segment, base=0x00000000, limit=0x00000000, Execute-Only, 16-bit
GDT[0x02]=LDT