[PATCH 4/4] tty: Use RCU read lock to iterate tasks in __do_SAK()

[Kirill Tkhai]

There were made several efforts to make __do_SAK()
working in process context long ago, but it does
not solves the problem completely. Since __do_SAK()
may take tasklist_lock for a long time, the concurent
processes waiting for write lock with interrupts
disabled (e.g., forking), get into the same situation
like __do_SAK() would have been executed in interrupt
context. I've observed several hard lockups on 3.10
kernel running 200 containers, caused by long duration
of copy_process()->write_lock_irq() after SAK was sent
to a tty. Current mainline kernel has the same problem.
So, this patch solves the problem and makes __do_SAK()
to be not greedy of tasklist_lock.

The solution is to use RCU to iterate processes and files.
Task list integrity is the only reason we taken tasklist_lock
before, as tty subsys primitives mostly take it for reading
also (e.g., __proc_set_tty). RCU read lock is enough for that.
Task list lock is taken for a small duration to check we've
iterated all the list and do not race with newly forked children.
It's not really useful, but we iterated the whole list before,
so let's save the old behaviour.

This patch covers all the cases, when __do_SAK() used
to find a tty-related task before. Also, I reordered
file iterations and p->signal->tty check and added
tty_lock() to close small race with the case, when
task obtains tty fd and then calls __proc_set_tty()
in parallel with its processing.

The patch is aimed to prevent hard lockups I've pointed above.

Signed-off-by: Kirill Tkhai <ktkhai@xxxxxxxxxxxxx>
---
 drivers/tty/tty_io.c |   28 ++++++++++++++++++++++------
 1 file changed, 22 insertions(+), 6 deletions(-)

diff --git a/drivers/tty/tty_io.c b/drivers/tty/tty_io.c
index 94813ae40983..2b6afe9e7bbb 100644
--- a/drivers/tty/tty_io.c
+++ b/drivers/tty/tty_io.c
@@ -2707,6 +2707,7 @@ void __do_SAK(struct tty_struct *tty)
 	struct task_struct *p;
 	struct pid *session;
 	int		i;
+	bool locked = false;
 
 	if (!tty)
 		return;
@@ -2723,15 +2724,12 @@ void __do_SAK(struct tty_struct *tty)
 			   task_pid_nr(p), p->comm);
 		send_sig(SIGKILL, p, 1);
 	} while_each_pid_task(session, PIDTYPE_SID, p);
+	read_unlock(&tasklist_lock);
 
+	tty_lock(tty);
+	rcu_read_lock();
 	/* Now kill any processes that happen to have the tty open */
 	for_each_process(p) {
-		if (p->signal->tty == tty) {
-			tty_notice(tty, "SAK: killed process %d (%s): by controlling tty\n",
-				   task_pid_nr(p), p->comm);
-			send_sig(SIGKILL, p, 1);
-			continue;
-		}
 		task_lock(p);
 		i = iterate_fd(p->files, 0, this_tty, tty);
 		if (i != 0) {
@@ -2740,8 +2738,26 @@ void __do_SAK(struct tty_struct *tty)
 			force_sig(SIGKILL, p);
 		}
 		task_unlock(p);
+
+		/*
+		 * p->signal is always valid for task_struct obtained
+		 * from the task list under rcu_read_lock().
+		 */
+		if (!i && p->signal->tty == tty) {
+			tty_notice(tty, "SAK: killed process %d (%s): by controlling tty\n",
+				   task_pid_nr(p), p->comm);
+			send_sig(SIGKILL, p, 1);
+		}
+
+		if (unlikely(next_task(p) == &init_task && !locked)) {
+			/* Take the lock to pick newly forked tasks */
+			read_lock(&tasklist_lock);
+			locked = true;
+		}
 	}
 	read_unlock(&tasklist_lock);
+	rcu_read_unlock();
+	tty_unlock(tty);
 #endif
 }