X-Git-Url: http://git.tuebingen.mpg.de/?a=blobdiff_plain;f=sched.c;fp=sched.c;h=6b8e09108308336676e48c4d8c9e84eb6e388431;hb=8bf35b38357c3ce59f52ae87f6e84e4b6d183ac7;hp=101cc4c0dad3706e841881b2a6d2989825fc58b0;hpb=a96be5bc4fda8c0df5370d646defb5ff632ba391;p=paraslash.git

diff --git a/sched.c b/sched.c
index 101cc4c0..6b8e0910 100644
--- a/sched.c
+++ b/sched.c
@@ -75,8 +75,11 @@ static unsigned sched_post_select(struct sched *s)
 	unsigned num_running_tasks = 0;
 
 	list_for_each_entry_safe(t, tmp, &s->task_list, node) {
-		if (t->error < 0)
+		if (t->error < 0) {
+			if (t->dead) /* task has been reaped */
+				unlink_and_free_task(t);
 			continue;
+		}
 		call_post_select(s, t);
 		t->notification = 0;
 		if (t->error < 0) {
@@ -138,6 +141,53 @@ again:
 	goto again;
 }
 
+/**
+ * Obtain the error status of a task and deallocate its resources.
+ *
+ * \param tptr Identifies the task to reap.
+ *
+ * This function is similar to wait(2) in that it returns information about a
+ * terminated task and allows to release the resources associated with the
+ * task. Until this function is called, the terminated task remains in a zombie
+ * state.
+ *
+ * \return If \a tptr is \p NULL, or \a *tptr is \p NULL, the function does
+ * nothing and returns zero. Otherwise, it is checked whether the task
+ * identified by \a tptr is still running. If it is, the function returns zero
+ * and again, no action is taken. Otherwise the (negative) error code of the
+ * terminated task is returned and \a *tptr is set to \p NULL. The task will
+ * then be removed removed from the scheduler task list.
+ *
+ * \sa \ref sched_shutdown(), wait(2).
+ */
+int task_reap(struct task **tptr)
+{
+	struct task *t;
+
+	if (!tptr)
+		return 0;
+	t = *tptr;
+	if (!t)
+		return 0;
+	if (!t->owned_by_sched)
+		return 0;
+	if (t->error >= 0)
+		return 0;
+	if (t->dead) /* will be freed in sched_post_select() */
+		return 0;
+	/*
+	 * With list_for_each_entry_safe() it is only safe to remove the
+	 * _current_ list item. Since we are being called from the loop in
+	 * schedule() via some task's ->post_select() function, freeing the
+	 * given task here would result in use-after-free bugs in schedule().
+	 * So we only set t->dead which tells schedule() to free the task in
+	 * the next iteration of its loop.
+	 */
+	t->dead = true;
+	*tptr = NULL;
+	return t->error;
+}
+
 /**
  * Deallocate all resources of all tasks of a scheduler instance.
  *
@@ -199,6 +249,7 @@ struct task *task_register(struct task_info *info, struct sched *s)
 	t->status[sizeof(t->status) - 1] = '\0';
 	t->notification = 0;
 	t->error = 0;
+	t->dead = false;
 	t->pre_select = info->pre_select;
 	t->post_select = info->post_select;
 	t->context = info->context;
@@ -239,7 +290,7 @@ char *get_task_list(struct sched *s)
 	list_for_each_entry_safe(t, tmp, &s->task_list, node) {
 		char *tmp_msg;
 		tmp_msg = make_message("%s%p\t%s\t%s\n", msg? msg : "", t,
-			t->error < 0? "zombie" : "running",
+			t->error < 0? (t->dead? "dead" : "zombie") : "running",
 			t->status);
 		free(msg);
 		msg = tmp_msg;