Hi Peter,

Em 04/11/2024 08:22, Peter Zijlstra escreveu:

On Fri, Nov 01, 2024 at 01:21:46PM -0300, André Almeida wrote:

quoted

@@ -1046,24 +1095,44 @@ static inline void exit_pi_state_list(struct task_struct *curr) { }
  
  static void futex_cleanup(struct task_struct *tsk)
  {
+	struct robust_list2_entry *curr, *n;
+	struct list_head *list2 = &tsk->robust_list2;
+
  #ifdef CONFIG_64BIT
  	if (unlikely(tsk->robust_list)) {
-		exit_robust_list64(tsk);
+		exit_robust_list64(tsk, tsk->robust_list);
  		tsk->robust_list = NULL;
  	}
  #else
  	if (unlikely(tsk->robust_list)) {
-		exit_robust_list32(tsk);
+		exit_robust_list32(tsk, (struct robust_list_head32 *) tsk->robust_list);
  		tsk->robust_list = NULL;
  	}
  #endif
  
  #ifdef CONFIG_COMPAT
  	if (unlikely(tsk->compat_robust_list)) {
-		exit_robust_list32(tsk);
+		exit_robust_list32(tsk, tsk->compat_robust_list);
  		tsk->compat_robust_list = NULL;
  	}
  #endif
+	/*
+	 * Walk through the linked list, parsing robust lists and freeing the
+	 * allocated lists
+	 */
+	if (unlikely(!list_empty(list2))) {
+		list_for_each_entry_safe(curr, n, list2, list) {
+			if (curr->head != NULL) {
+				if (curr->list_type == ROBUST_LIST_64BIT)
+					exit_robust_list64(tsk, curr->head);
+				else if (curr->list_type == ROBUST_LIST_32BIT)
+					exit_robust_list32(tsk, curr->head);
+				curr->head = NULL;
+			}
+			list_del_init(&curr->list);
+			kfree(curr);
+		}
+	}
  
  	if (unlikely(!list_empty(&tsk->pi_state_list)))
  		exit_pi_state_list(tsk);

I'm still digesting this, but the above seems particularly silly.

Should not the legacy lists also be on the list of lists? I mean, it
makes no sense to have two completely separate means of tracking lists.

You are asking if, whenever someone calls set_robust_list() or 
compat_set_robust_list() to be inserted into &current->robust_list2 
instead of using tsk->robust_list and tsk->compat_robust_list?

I was thinking of doing that, but my current implementation has a 
kmalloc() call for every insertion, and I wasn't sure if I could add 
this new latency to the old set_robust_list() syscall. Assuming it is 
usually called just once during the thread initialization perhaps it 
shouldn't cause much harm I guess.