diff --git a/src/thread/thread.c b/src/thread/thread.c
index 4c9f740..fa7ef59 100644
--- a/src/thread/thread.c
+++ b/src/thread/thread.c
@@ -12,16 +12,16 @@
 
 #define FINISHED  0x1
 #define ALLOCATED 0x2
-#define WAITING   0x4
 
 #ifndef STACK_SIZE
-#define STACK_SIZE 4096
+#define STACK_SIZE 4096*4
 #endif
 
 struct context_entry {
   TAILQ_ENTRY(context_entry) link;
   ucontext_t context;
   thread_t id;
+  thread_t parent;
   void *retvalue;
   int valgrind_id;
   char status;
@@ -32,11 +32,10 @@ static struct context_entry *running = 0;
 static unsigned long long counter = 0;
 
 int thread_yield(void) {
-  TRACE("thread_yield");
   if (counter <= 1) {
     return 0;
   }
-  struct context_entry *first = NULL;
+  struct context_entry *first;
   int count = 0;
   do  {
     if (count++ == counter) {
@@ -51,9 +50,7 @@ int thread_yield(void) {
     if (first->id == running->id) {
       return 0;
     }
-  } while ((first->status & FINISHED) ||
-    ((first->status & WAITING) && !(((struct context_entry *)first->retvalue)->status & FINISHED)));
-  TRACE("PICKING %p", first);
+  } while (first->status & FINISHED);
   struct context_entry *old_runner = running;
   running = first;
   swapcontext(&old_runner->context, &running->context);
@@ -67,25 +64,21 @@ thread_t thread_self(void) {
   return running->id;
 }
 
-void thread_function_wrapper(void *(*func)(void *), void *funcarg) {
-  TRACE("Wrapper for %p\n", func);
-  thread_exit(func(funcarg));
-}
-
 int thread_create(thread_t *newthread, void *(*func)(void *), void *funcarg) {
   TRACE("Create a new thread that execute function %p", func);
-  struct context_entry *new_entry = malloc(sizeof(*new_entry));
   if (counter == 0) {
-    memset(new_entry, 0, sizeof(*new_entry));
-    getcontext(&new_entry->context);
-    new_entry->id = 0;
-    new_entry->status = 0;
-    new_entry->retvalue = 0;
-    TAILQ_INSERT_HEAD(&head, new_entry, link);
-    running = new_entry;
+    struct context_entry *main = malloc(sizeof(*main));
+    memset(main, 0, sizeof(*main));
+    getcontext(&main->context);
+    main->id = 0;
+    main->parent = 0;
+    main->status = 0;
+    main->retvalue = 0;
+    TAILQ_INSERT_HEAD(&head, main, link);
+    running = main;
     counter++;
-    new_entry = malloc(sizeof(*new_entry));
   }
+  struct context_entry *new_entry = malloc(sizeof(*new_entry));
   memset(new_entry, 0, sizeof(*new_entry));
   getcontext(&new_entry->context);
   new_entry->context.uc_stack.ss_sp = malloc(STACK_SIZE);
@@ -95,17 +88,18 @@ int thread_create(thread_t *newthread, void *(*func)(void *), void *funcarg) {
     new_entry->context.uc_stack.ss_sp + new_entry->context.uc_stack.ss_size
     );
   new_entry->id = new_entry;
+  new_entry->parent = running->id;
   new_entry->status = ALLOCATED;
   new_entry->retvalue = 0;
   *newthread = new_entry;
-  makecontext(&new_entry->context, (void (*)(void)) thread_function_wrapper, 2, func, funcarg);
+  makecontext(&new_entry->context, (void (*)()) func, 1, funcarg);
   counter++;
   TAILQ_INSERT_HEAD(&head, new_entry, link);
   return 0;
 }
 
 void print_entry(struct context_entry *entry) {
-  TRACE("CONTEXT (%p, %p, %d);", entry, entry->id, (entry->status & FINISHED) > 0);
+  TRACE("CONTEXT (%p, %p, %d, %p);\n", entry, entry->id, (entry->status & FINISHED) > 0, entry->parent);
 }
 
 int thread_join(thread_t thread, void **retval) {
@@ -115,14 +109,12 @@ int thread_join(thread_t thread, void **retval) {
     if (entry->id != thread) {
       continue;
     }
-    TRACE("FIND %d",entry->status);
-    running->status |= WAITING;
-    running->retvalue = entry;
-    while (!(entry->status & FINISHED)) {
+    TRACE("FIND");
+    print_entry(entry);
+    while (!entry->status & FINISHED) {
       TRACE("NOT FINISHED");
       thread_yield();
     }
-    running->status &= ~WAITING;
     TRACE("AFTER");
     if (retval)
       *retval = entry->retvalue;
@@ -152,8 +144,6 @@ int thread_join(thread_t thread, void **retval) {
 
 void thread_exit(void *retval) {
   TRACE("Exit thread %p", running);
-  if (running == 0)
-    exit(0);
   running->status |= FINISHED;
   running->retvalue = retval;
   thread_yield();