Linus spotted one cut-n-pasto ('tracing' argument) but didn't see the
other: we were walking the ptrace_children list by the sibling field.

So we got garbage for your task_structs when this happened.  If the list
wasn't empty, it would crash.  Strace detaches from all tasks when it
receives a Control-C so only with enough threads and SMP would this be
easily seen.