diff --git a/src/kernel/boot/kmain.c b/src/kernel/boot/kmain.c
index fe2b320..bbd10a8 100644
--- a/src/kernel/boot/kmain.c
+++ b/src/kernel/boot/kmain.c
@@ -8,16 +8,9 @@
 #include <process.h>
 #include <scheduler.h>
 
-int *pid = (int *)0x20000;
 void thread_function()
 {
-  *pid = process()->pid;
-
-  while(1)
-  {
-    debug("Process %d\n", *pid);
-    yield();
-  }
+  while(1);
 }
 
 
@@ -35,16 +28,12 @@ void kmain(uint64_t multiboot_magic, void *multiboot_data)
 
   cpu_init();
 
-  struct process *p1 = new_process(thread_function);
-  vmm_set_page(p1->P4, 0x20000, pmm_alloc(), PAGE_WRITE | PAGE_PRESENT);
-  struct process *p2 = new_process(thread_function);
-  vmm_set_page(p2->P4, 0x20000, pmm_alloc(), PAGE_WRITE | PAGE_PRESENT);
-  struct process *p3 = new_process(thread_function);
-  vmm_set_page(p3->P4, 0x20000, pmm_alloc(), PAGE_WRITE | PAGE_PRESENT);
+  struct process *p1 = new_process((void (*)(void))0x10000);
+  uint64_t page = pmm_alloc();
+  vmm_set_page(p1->P4, 0x10000, page, PAGE_WRITE | PAGE_PRESENT | PAGE_USER);
+  memcpy(P2V(page), (void *)(uintptr_t)thread_function, PAGE_SIZE);
 
   ready(p1);
-  ready(p2);
-  ready(p3);
 
   debug_ok("Boot \"Complete\"\n");
 
diff --git a/src/kernel/cpu/cpu.c b/src/kernel/cpu/cpu.c
index 1970cae..da598c5 100644
--- a/src/kernel/cpu/cpu.c
+++ b/src/kernel/cpu/cpu.c
@@ -1,7 +1,7 @@
 #include <cpu.h>
 #include <interrupts.h>
 
-uint64_t gdt[5];
+uint64_t gdt[6];
 uint8_t tss[104];
 void gdt_init(uint64_t *, void *);
 
diff --git a/src/kernel/cpu/gdt.c b/src/kernel/cpu/gdt.c
index ed89f7f..7962337 100644
--- a/src/kernel/cpu/gdt.c
+++ b/src/kernel/cpu/gdt.c
@@ -6,7 +6,7 @@
 #define GDT_DPL(lvl)  ((lvl)<<13)
 #define GDT_PRESENT   (1<<15)
 #define GDT_LONG      (1<<21)
-#define GDT_TSS       (1<<8)
+#define GDT_TSS       (9<<8)
 
 struct gdt
 {
@@ -44,6 +44,7 @@ struct gdt BootGDT[] = {
   {0, 0},
   {0, GDT_PRESENT | GDT_DPL(0) | GDT_CODE | GDT_LONG},
   {0, GDT_PRESENT | GDT_DPL(3) | GDT_CODE | GDT_LONG},
+  {0, GDT_PRESENT | GDT_DPL(3) | (1<<12) | (1<<9)},
   {0, 0},
   {0, 0},
 };
@@ -56,18 +57,27 @@ void gdt_init(uint64_t *_gdt, void *_tss)
   struct gdt *gdt = (struct gdt *)_gdt;
   memcpy(gdt, BootGDT, sizeof(BootGDT));
 
+  struct tss *tss = _tss;
+  tss->io_mba = sizeof(struct tss);
+
   uint32_t tss_limit = sizeof(struct tss);
   uint64_t tss_base = (uint64_t)_tss;
-  gdt[3].flags = GDT_PRESENT | GDT_TSS;
-  gdt[3].flags |= (((tss_base >> 24) & 0xFF) << 24);
-  gdt[3].flags |= ((tss_base >> 16) & 0xFF);
-  gdt[3].flags |= (((tss_limit >> 16) & 0xF) << 16);
-  gdt[3].addr = ((tss_base & 0xFFFF) << 16) | (tss_limit & 0xFFFF);
-  gdt[4].flags = 0;
-  gdt[4].addr = ((tss_base >> 32) & 0xFFFF);
+  gdt[4].flags = GDT_PRESENT | GDT_TSS;
+  gdt[4].flags |= (((tss_base >> 24) & 0xFF) << 24);
+  gdt[4].flags |= ((tss_base >> 16) & 0xFF);
+  gdt[4].flags |= (((tss_limit >> 16) & 0xF) << 16);
+  gdt[4].addr = ((tss_base & 0xFFFF) << 16) | (tss_limit & 0xFFFF);
+  gdt[5].flags = 0;
+  gdt[5].addr = ((tss_base >> 32) & 0xFFFFFFFF);
 
-  GDTp.len = 5*8 - 1;
+  GDTp.len = 6*8 - 1;
   GDTp.gdt = gdt;
 
   load_gdt(&GDTp);
 }
+
+void set_tss_rsp0(void *_tss, void *rsp0)
+{
+  struct tss *tss = _tss;
+  tss->rsp0 = (uint64_t)rsp0;
+}
diff --git a/src/kernel/cpu/registers.S b/src/kernel/cpu/registers.S
index 6d75804..f99166d 100644
--- a/src/kernel/cpu/registers.S
+++ b/src/kernel/cpu/registers.S
@@ -17,7 +17,7 @@ load_gdt:
   push rax
   retfq
 .load_gdt:
-  mov rax, 0x18
+  mov rax, 0x20
   ltr ax
   ret
 
diff --git a/src/kernel/include/cpu.h b/src/kernel/include/cpu.h
index 9cf84a4..beba505 100644
--- a/src/kernel/include/cpu.h
+++ b/src/kernel/include/cpu.h
@@ -3,6 +3,8 @@
 
 void cpu_init();
 
+extern uint8_t tss[];
+void set_tss_rsp0(void *tss, void *rsp0);
 
 
 void load_idt(void *);
diff --git a/src/kernel/include/memory.h b/src/kernel/include/memory.h
index e4ae958..13e8035 100644
--- a/src/kernel/include/memory.h
+++ b/src/kernel/include/memory.h
@@ -18,6 +18,7 @@
 
 #define PAGE_PRESENT      0x001
 #define PAGE_WRITE        0x002
+#define PAGE_USER         0x004
 #define PAGE_HUGE         0x080
 #define PAGE_GLOBAL       0x100
 
diff --git a/src/kernel/include/process.h b/src/kernel/include/process.h
index c375907..5473bb6 100644
--- a/src/kernel/include/process.h
+++ b/src/kernel/include/process.h
@@ -2,6 +2,7 @@
 #include <queue.h>
 #include <scheduler.h>
 #include <stdint.h>
+#include <interrupts.h>
 
 struct process
 {
diff --git a/src/kernel/proc/process.c b/src/kernel/proc/process.c
index ff2b57d..a52ea18 100644
--- a/src/kernel/proc/process.c
+++ b/src/kernel/proc/process.c
@@ -11,8 +11,10 @@ struct swtch_stack
   uint64_t R13;
   uint64_t R14;
   uint64_t R15;
+  uint64_t isr_return_arg;
   uint64_t RBP2;
   uint64_t ret;
+  registers r;
 }__attribute__((packed));
 
 struct process *sched_proc = 0;
@@ -28,7 +30,15 @@ struct process *new_process(void (*function)(void))
 
   struct swtch_stack *stk = proc->stack_ptr;
   stk->RBP = (uint64_t)&stk->RBP2;
-  stk->ret = (uint64_t)function;
+
+  stk->ret = (uint64_t)isr_return;
+  stk->isr_return_arg = (uint64_t)&stk->r;
+
+  stk->r.rip = (uint64_t)function;
+  stk->r.cs = 0x10 | 3;
+  stk->r.ss = 0x18 | 3;
+  stk->r.rflags = 3<<12;
+  stk->r.rsp = 0x10FFF;
 
   return proc;
 }
@@ -52,6 +62,7 @@ void scheduler()
 
     _proc = new;
     write_cr3(new->P4);
+    set_tss_rsp0(tss, new + PAGE_SIZE);
     switch_stack(&sched_proc->stack_ptr, &new->stack_ptr);
 
     ready(_proc);
@@ -61,8 +72,15 @@ void scheduler()
 
 void start_scheduler()
 {
-  sched_proc = new_process(scheduler);
+  sched_proc = P2V(pmm_calloc());
   sched_proc->pid = (uint64_t)-1;
+  sched_proc->stack_ptr = incptr(sched_proc, PAGE_SIZE - sizeof(struct swtch_stack) + sizeof(registers));
+  sched_proc->P4 = kernel_P4;
+
+  struct swtch_stack *stk = sched_proc->stack_ptr;
+  stk->RBP = (uint64_t)&stk->RBP2;
+
+  stk->ret = (uint64_t)scheduler;
 
   uint64_t stack;
   switch_stack(&stack, &sched_proc->stack_ptr);
diff --git a/src/kernel/proc/swtch.S b/src/kernel/proc/swtch.S
index 097e8ec..5ee6574 100644
--- a/src/kernel/proc/swtch.S
+++ b/src/kernel/proc/swtch.S
@@ -5,6 +5,7 @@
 switch_stack:
   push rbp
   mov rbp, rsp
+  push rdi
   push r15
   push r14
   push r13
@@ -21,5 +22,6 @@ switch_stack:
   pop r13
   pop r14
   pop r15
+  pop rdi
   leaveq
   ret