Back about a year ago when I last fiddled heavily with the do_wait code, I
was thinking too hard about the wrong thing and I now think I introduced a
bug whose inverse thought I was fixing. Apparently noone was looking too
hard over much shoulder, so as to cite my bogus reasoning at the time. In
the race condition when PTRACE_ATTACH is about to steal a child and then
the child hits a tracing event (what my_ptrace_child checks for), the real
parent does need to set its flag noting it has some eligible live children.
Otherwise a spurious ECHILD error is possible, since the child in question
is not yet on the ptrace_children list.
Signed-off-by: Roland McGrath <[email protected]>
--- linux-2.6/kernel/exit.c
+++ linux-2.6/kernel/exit.c
@@ -1370,6 +1370,15 @@ repeat:
switch (p->state) {
case TASK_TRACED:
+ /*
+ * When we hit the race with PTRACE_ATTACH,
+ * we will not report this child. But the
+ * race means it has not yet been moved to
+ * our ptrace_children list, so we need to
+ * set the flag here to avoid a spurious ECHILD
+ * when the race happens with the only child.
+ */
+ flag = 1;
if (!my_ptrace_child(p))
continue;
/*FALLTHROUGH*/