On Fri, Jul 12, 2024 at 02:28:13PM -0700, Andrii Nakryiko wrote:

quoted

@@ -1814,7 +1822,7 @@ static int dup_utask(struct task_struct
                        return -ENOMEM;

                *n = *o;
-               get_uprobe(n->uprobe);
+               __srcu_clone_read_lock(&uretprobes_srcu, n->srcu_idx);

do we need to add this __srcu_clone_read_lock hack just to avoid
taking a refcount in dup_utask (i.e., on process fork)? This is not
that frequent and performance-sensitive case, so it seems like it
should be fine to take refcount and avoid doing srcu_read_unlock() in
a new process. Just like the case with long-running uretprobes where
you convert SRCU lock into refcount.

Yes, I suppose that is now possible too. But it makes the patches harder
to split. Let me ponder that after I get it to pass your stress thing.