#include <libk/std.h>
#include <libk/string.h>
#include <amd64/init.h>
#include <amd64/tss.h>
#include <amd64/debug.h>

#define GDT_KCODE     0x08
#define GDT_KDATA     0x10
#define GDT_UCODE     0x18
#define GDT_UDATA     0x20
#define GDT_TSS       0x28

#define TSS         0x80
#define TSS_PRESENT 0x89

#define KSTACK_SIZE   (8*1024)

struct gdt_entry {
  uint16_t limitlow;
  uint16_t baselow;
  uint8_t basemid;
  uint8_t access;
  uint8_t gran;
  uint8_t basehigh;
} __attribute__((packed));

struct gdt_ptr {
  uint16_t limit;
  uint64_t base;
} __attribute__((packed));

struct gdt_extended {
  struct gdt_entry old[5];
  struct gdt_entry tsslow;
  struct gdt_entry tsshigh;
} __attribute__((packed));

__attribute__((aligned(16))) static volatile uint8_t kernel_stack[KSTACK_SIZE];
__attribute__((aligned(16))) static volatile struct gdt_extended gdt;

static void amd64_gdt_set(volatile struct gdt_entry *ent, uint32_t base,
                          uint32_t limit, uint8_t acc, uint8_t gran) {
  ent->baselow = (base & 0xFFFF);
  ent->basemid = (base >> 16) & 0xFF;
  ent->basehigh = (base >> 24) & 0xFF;
  ent->limitlow = (limit & 0xFFFF);
  ent->gran = ((limit >> 16) & 0x0F) | (gran & 0xF0);
  ent->access = acc;
}

static void amd64_gdt_init(void) {
  volatile struct tss *tss = amd64_get_tss();

  memset((void *)&gdt, 0, sizeof(gdt));
  memset((void *)kernel_stack, 0, sizeof(kernel_stack));
  memset((void *)tss, 0, sizeof(*tss));

  tss->iopb_off = sizeof(*tss);
  tss->rsp0 = (uint64_t)((uintptr_t)kernel_stack + sizeof(kernel_stack));

  uint64_t tssbase = (uint64_t)&tss;
  uint64_t tsslimit = sizeof(*tss) - 1;

  amd64_gdt_set(&gdt.old[0], 0, 0, 0, 0);
  amd64_gdt_set(&gdt.old[1], 0, 0xFFFFF, 0x9A, 0xA0);
  amd64_gdt_set(&gdt.old[2], 0, 0xFFFFF, 0x92, 0xC0);
  amd64_gdt_set(&gdt.old[3], 0, 0xFFFFF, 0xFA, 0xA0);
  amd64_gdt_set(&gdt.old[4], 0, 0xFFFFF, 0xF2, 0xC0);
  amd64_gdt_set(&gdt.tsslow, (tssbase & 0xFFFFFFFF), tsslimit, TSS_PRESENT | TSS, 0);

  uint32_t tssbasehigh = (tssbase >> 32);
  gdt.tsshigh.limitlow = (tssbasehigh & 0xFFFF);
  gdt.tsshigh.baselow = (tssbasehigh >> 16) & 0xFFFF;
  gdt.tsshigh.basemid = 0;
  gdt.tsshigh.basehigh = 0;
  gdt.tsshigh.access = 0;
  gdt.tsshigh.gran = 0;

  struct gdt_ptr gdtr;
  gdtr.limit = sizeof(gdt) - 1;
  gdtr.base = (uint64_t)&gdt;
  __asm__ volatile("lgdt %0" :: "m"(gdtr) : "memory");

  __asm__ volatile(
    "pushq %[kcode]\n"
    "lea 1f(%%rip), %%rax\n"
    "pushq %%rax\n"
    "lretq\n"
    "1:\n"
    "movw %[kdata], %%ax\n"
    "movw %%ax, %%ds\n"
    "movw %%ax, %%es\n"
    "movw %%ax, %%ss\n"
    :
    : [kcode] "i"(GDT_KCODE), [kdata] "i"(GDT_KDATA)
    : "rax", "memory"
  );

  __asm__ volatile("ltr %0" :: "r"((uint16_t)GDT_TSS));
}

void amd64_init(void) {
  amd64_gdt_init();
  amd64_debug_init();
}