#include <amd64/apic.h>
#include <amd64/init.h>
#include <amd64/intr_defs.h>
#include <amd64/mm.h>
#include <amd64/msr-index.h>
#include <amd64/msr.h>
#include <libk/std.h>
#include <libk/string.h>
#include <limine/requests.h>
#include <mm/liballoc.h>
#include <proc/proc.h>
#include <sync/spin_lock.h>
#include <sys/debug.h>
#include <sys/sched.h>
#include <sys/smp.h>
#include <sys/syscall.h>

/// The CPUs
static struct cpu cpus[CPUS_MAX];

static atomic_int last_cpu_index = 0;
static atomic_int cpu_counter;
static atomic_int cpu_id_counter = 0;

/// Allocate a CPU structure
struct cpu* cpu_make (uint64_t lapic_id, uint64_t acpi_id) {
  int id = atomic_fetch_add (&cpu_id_counter, 1);
  struct cpu* cpu = &cpus[id];

  memset (cpu, 0, sizeof (*cpu));
  cpu->lock = SPIN_LOCK_INIT;
  cpu->id = id;
  cpu->acpi_id = acpi_id;
  cpu->lapic_id = lapic_id;

  amd64_wrmsr (MSR_GS_BASE, (uint64_t)cpu);

  return cpu;
}

struct cpu* cpu_get (void) {
  struct cpu* ptr = (struct cpu*)amd64_rdmsr (MSR_GS_BASE);
  return ptr;
}

void cpu_request_sched (struct cpu* cpu) {
  if (cpu == thiscpu) {
    proc_sched ();
    return;
  }

  amd64_lapic_ipi (cpu->lapic_id, CPU_REQUEST_SCHED);
}

struct cpu* cpu_find_lightest (void) {
  struct limine_mp_response* mp = limine_mp_request.response;

  int start = atomic_fetch_add (&last_cpu_index, 1) % mp->cpu_count;
  struct cpu* best_cpu = &cpus[start];
  int best_load = atomic_load (&best_cpu->proc_run_q_count);

  for (int i = 1; i < (int)mp->cpu_count; i++) {
    int idx = (start + i) % mp->cpu_count;
    struct cpu* cpu = &cpus[idx];
    int l = atomic_load (&cpu->proc_run_q_count);
    if (l < best_load) {
      best_load = l;
      best_cpu = cpu;
    }
  }

  return best_cpu;
}

/// Bootstrap code for non-BSP CPUs
static void amd64_smp_bootstrap (struct limine_mp_info* mp_info) {
  amd64_load_kernel_cr3 ();

  struct cpu* cpu = cpu_make (mp_info->lapic_id, mp_info->processor_id);

  amd64_init (cpu, true); /* gdt + idt */
  syscall_init ();

  amd64_lapic_init (1000);

  DEBUG ("CPU %u is online!\n", thiscpu->id);

  atomic_fetch_sub (&cpu_counter, 1);

  struct proc* spin_proc = proc_spawn_rd ("spin.exe");
  struct cpu* spin_cpu = thiscpu;
  proc_register (spin_proc, &spin_cpu);

  spin_lock_ctx_t ctxcpu;
  spin_lock (&spin_proc->cpu->lock, &ctxcpu);
  do_sched (spin_proc, &spin_proc->cpu->lock, &ctxcpu);
}

/// Initialize SMP subsystem for AMD64. Start AP CPUs
void smp_init (void) {
  amd64_lapic_init (1000);

  struct limine_mp_response* mp = limine_mp_request.response;

  cpu_counter = mp->cpu_count - 1;

  for (size_t i = 0; i < mp->cpu_count; i++) {
    if (mp->cpus[i]->processor_id != thiscpu->acpi_id) {
      mp->cpus[i]->goto_address = &amd64_smp_bootstrap;
    }
  }

  DEBUG ("Waiting for other CPUs:\n");

  while (atomic_load (&cpu_counter) > 0)
    debugprintf (".\n");

  DEBUG ("All CPUs are up!\n");
}