Skip to contentSkip to conversation

audit: Fixes and tasklist_lock -> RCU conversion - Resend

PATCH In review Thomas Gleixner 2 participants 7 messages Sep 7, 2010 – Sep 8, 2010

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

cover audit: Fixes and tasklist_lock -> RCU conversion - Resend

Hi,

the following patches result from auditing read_lock(&tasklist_lock) sites whether they can be converted to RCU.

Resend of http://lkml.org/lkml/2009/12/9/216

Thanks,

tglx

#1 lore raw

PATCH 1/3 audit: Call tty_audit_push_task() outside preempt disabled

While auditing all tasklist_lock read_lock sites I stumbled over the following call chain:

audit_prepare_user_tty()
read_lock(&tasklist_lock);

  tty_audit_push_task();
     mutex_lock(&buf->mutex);

     --> buf->mutex is locked with preemption disabled.

Solve this by acquiring a reference to the task struct under rcu_read_lock and call tty_audit_push_task outside of the preempt disabled region.

Move all code which needs to be protected by sighand lock into tty_audit_push_task() and use lock/unlock_sighand as we do not hold tasklist_lock.

Cc Al Viro <[email protected]>
Cc Eric Paris <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Oleg Nesterov <[email protected]>

---
drivers/char/tty_audit.c | 38 ++++++++++++++++++++++++++++----------

 drivers/char/tty_audit.c |   38 ++++++++++++++++++++++++++++----------
 include/linux/tty.h      |    8 ++------
 kernel/audit.c           |   25 +++++++++----------------
 3 files changed, 39 insertions(+), 32 deletions(-)
drivers/char/tty_audit.c +28 -10
Index: linux-2.6/drivers/char/tty_audit.c
===================================================================
@@ -188,25 +188,43 @@ void tty_audit_tiocsti(struct tty_struct
 }
 
 /**
- *	tty_audit_push_task	-	Flush task's pending audit data
+ * tty_audit_push_task	-	Flush task's pending audit data
+ * @tsk:		task pointer
+ * @loginuid:		sender login uid
+ * @sessionid:		sender session id
+ *
+ * Called with a ref on @tsk held. Try to lock sighand and get a
+ * reference to the tty audit buffer if available.
+ * Flush the buffer or return an appropriate error code.
  */
-void tty_audit_push_task(struct task_struct *tsk, uid_t loginuid, u32 sessionid)
+int tty_audit_push_task(struct task_struct *tsk, uid_t loginuid, u32 sessionid)
 {
-	struct tty_audit_buf *buf;
+	struct tty_audit_buf *buf = ERR_PTR(-EPERM);
+	unsigned long flags;
 
-	spin_lock_irq(&tsk->sighand->siglock);
-	buf = tsk->signal->tty_audit_buf;
-	if (buf)
-		atomic_inc(&buf->count);
-	spin_unlock_irq(&tsk->sighand->siglock);
-	if (!buf)
-		return;
+	if (!lock_task_sighand(tsk, &flags))
+		return -ESRCH;
+
+	if (tsk->signal->audit_tty) {
+		buf = tsk->signal->tty_audit_buf;
+		if (buf)
+			atomic_inc(&buf->count);
+	}
+	unlock_task_sighand(tsk, &flags);
+
+	/*
+	 * Return 0 when signal->audit_tty set
+	 * but tsk->signal->tty_audit_buf == NULL.
+	 */
+	if (!buf || IS_ERR(buf))
+		return PTR_ERR(buf);
 
 	mutex_lock(&buf->mutex);
 	tty_audit_buf_push(tsk, loginuid, sessionid, buf);
 	mutex_unlock(&buf->mutex);
 
 	tty_audit_buf_put(buf);
+	return 0;
 }
 
 /**
include/linux/tty.h +2 -6
Index: linux-2.6/include/linux/tty.h
===================================================================
@@ -539,8 +539,8 @@ extern void tty_audit_exit(void);
 extern void tty_audit_fork(struct signal_struct *sig);
 extern void tty_audit_tiocsti(struct tty_struct *tty, char ch);
 extern void tty_audit_push(struct tty_struct *tty);
-extern void tty_audit_push_task(struct task_struct *tsk,
-					uid_t loginuid, u32 sessionid);
+extern int tty_audit_push_task(struct task_struct *tsk,
+			       uid_t loginuid, u32 sessionid);
 #else
 static inline void tty_audit_add_data(struct tty_struct *tty,
 				      unsigned char *data, size_t size)
@@ -558,10 +558,6 @@ static inline void tty_audit_fork(struct
 static inline void tty_audit_push(struct tty_struct *tty)
 {
 }
-static inline void tty_audit_push_task(struct task_struct *tsk,
-					uid_t loginuid, u32 sessionid)
-{
-}
 #endif
 
 /* tty_io.c */
kernel/audit.c +9 -16
Index: linux-2.6/kernel/audit.c
===================================================================
@@ -467,23 +467,16 @@ static int audit_prepare_user_tty(pid_t 
 	struct task_struct *tsk;
 	int err;
 
-	read_lock(&tasklist_lock);
+	rcu_read_lock();
 	tsk = find_task_by_vpid(pid);
-	err = -ESRCH;
-	if (!tsk)
-		goto out;
-	err = 0;
-
-	spin_lock_irq(&tsk->sighand->siglock);
-	if (!tsk->signal->audit_tty)
-		err = -EPERM;
-	spin_unlock_irq(&tsk->sighand->siglock);
-	if (err)
-		goto out;
-
-	tty_audit_push_task(tsk, loginuid, sessionid);
-out:
-	read_unlock(&tasklist_lock);
+	if (!tsk) {
+		rcu_read_unlock();
+		return -ESRCH;
+	}
+	get_task_struct(tsk);
+	rcu_read_unlock();
+	err = tty_audit_push_task(tsk, loginuid, sessionid);
+	put_task_struct(tsk);
 	return err;
 }
 
#2 lore raw

PATCH 2/3 audit: Do not send uninitialized data for AUDIT_TTY_GET

In reply to Thomas Gleixner

audit_receive_msg() sends uninitialized data for AUDIT_TTY_GET when the task was not found.

Send reply only when task was found.

Cc Al Viro <[email protected]>
Cc Eric Paris <[email protected]>
Cc Peter Zijlstra <[email protected]>

---
kernel/audit.c | 6 ++++--

 kernel/audit.c |    6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)
kernel/audit.c +4 -2
Index: linux-2.6/kernel/audit.c
===================================================================
@@ -884,8 +884,10 @@ static int audit_receive_msg(struct sk_b
 			spin_unlock_irq(&tsk->sighand->siglock);
 		}
 		read_unlock(&tasklist_lock);
-		audit_send_reply(NETLINK_CB(skb).pid, seq, AUDIT_TTY_GET, 0, 0,
-				 &s, sizeof(s));
+
+		if (!err)
+			audit_send_reply(NETLINK_CB(skb).pid, seq,
+					 AUDIT_TTY_GET, 0, 0, &s, sizeof(s));
 		break;
 	}
 	case AUDIT_TTY_SET: {
#3 lore raw

PATCH 3/3 audit: Use rcu for task lookup protection

In reply to Thomas Gleixner

Protect the task lookups in audit_receive_msg() with rcu_read_lock() instead of tasklist_lock and use lock/unlock_sighand to protect against the exit race.

Cc Al Viro <[email protected]>
Cc Eric Paris <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Oleg Nesterov <[email protected]>

---
kernel/audit.c | 30 ++++++++++++++----------------

 kernel/audit.c |   30 ++++++++++++++----------------
 1 file changed, 14 insertions(+), 16 deletions(-)
kernel/audit.c +14 -16
Index: linux-2.6/kernel/audit.c
===================================================================
@@ -873,17 +873,16 @@ static int audit_receive_msg(struct sk_b
 	case AUDIT_TTY_GET: {
 		struct audit_tty_status s;
 		struct task_struct *tsk;
+		unsigned long flags;
 
-		read_lock(&tasklist_lock);
+		rcu_read_lock();
 		tsk = find_task_by_vpid(pid);
-		if (!tsk)
-			err = -ESRCH;
-		else {
-			spin_lock_irq(&tsk->sighand->siglock);
+		if (tsk && lock_task_sighand(tsk, &flags)) {
 			s.enabled = tsk->signal->audit_tty != 0;
-			spin_unlock_irq(&tsk->sighand->siglock);
-		}
-		read_unlock(&tasklist_lock);
+			unlock_task_sighand(tsk, &flags);
+		} else
+			err = -ESRCH;
+		rcu_read_unlock();
 
 		if (!err)
 			audit_send_reply(NETLINK_CB(skb).pid, seq,
@@ -893,22 +892,21 @@ static int audit_receive_msg(struct sk_b
 	case AUDIT_TTY_SET: {
 		struct audit_tty_status *s;
 		struct task_struct *tsk;
+		unsigned long flags;
 
 		if (nlh->nlmsg_len < sizeof(struct audit_tty_status))
 			return -EINVAL;
 		s = data;
 		if (s->enabled != 0 && s->enabled != 1)
 			return -EINVAL;
-		read_lock(&tasklist_lock);
+		rcu_read_lock();
 		tsk = find_task_by_vpid(pid);
-		if (!tsk)
-			err = -ESRCH;
-		else {
-			spin_lock_irq(&tsk->sighand->siglock);
+		if (tsk && lock_task_sighand(tsk, &flags)) {
 			tsk->signal->audit_tty = s->enabled != 0;
-			spin_unlock_irq(&tsk->sighand->siglock);
-		}
-		read_unlock(&tasklist_lock);
+			unlock_task_sighand(tsk, &flags);
+		} else
+			err = -ESRCH;
+		rcu_read_unlock();
 		break;
 	}
 	default:
#4 lore raw

Re: [patch 3/3] audit: Use rcu for task lookup protection

Quoted 4 lines · Thomas Gleixner

Protect the task lookups in audit_receive_msg() with rcu_read_lock() instead of tasklist_lock and use lock/unlock_sighand to protect against the exit race.

I do not understand audit, but I belive both 1/3 and 3/3 patches are fine (I didn't get 2/3).

But, sorry, can't resists ;) off-topic nit.

Quoted 5 lines
kernel/audit.c +1 -0
@@ -873,17 +873,16 @@ static int audit_receive_msg(struct sk_b
 	case AUDIT_TTY_GET: {
 		struct audit_tty_status s;
 		struct task_struct *tsk;
+		unsigned long flags;

kernel/audit.c static int audit_receive_msg(struct sk_b · line 881

Quoted 10 lines
kernel/audit.c +2 -5
 
-		read_lock(&tasklist_lock);
+		rcu_read_lock();
 		tsk = find_task_by_vpid(pid);
-		if (!tsk)
-			err = -ESRCH;
-		else {
-			spin_lock_irq(&tsk->sighand->siglock);
+		if (tsk && lock_task_sighand(tsk, &flags)) {
 			s.enabled = tsk->signal->audit_tty != 0;

Yes, this is what original code does, it takes ->siglock every time around read/write of ->audit_tty. And this looks absolutely bogus. Say, tty_audit_fork(). Why does it take ->siglock ?

As for ->tty_audit_buf, I am not sure ->siglock is the best choice, perhaps task_lock() would be better.

Once again, I think the patch is fine. Just it seems to me this code needs more cleanups.

Oleg.

#5 lore raw

Re: [patch 3/3] audit: Use rcu for task lookup protection

Quoted 33 lines · Oleg Nesterov
Quoted 4 lines · Thomas Gleixner

Protect the task lookups in audit_receive_msg() with rcu_read_lock() instead of tasklist_lock and use lock/unlock_sighand to protect against the exit race.

I do not understand audit, but I belive both 1/3 and 3/3 patches are fine (I didn't get 2/3).

But, sorry, can't resists ;) off-topic nit.

Quoted 5 lines
kernel/audit.c +1 -0
@@ -873,17 +873,16 @@ static int audit_receive_msg(struct sk_b
 	case AUDIT_TTY_GET: {
 		struct audit_tty_status s;
 		struct task_struct *tsk;
+		unsigned long flags;

kernel/audit.c static int audit_receive_msg(struct sk_b · line 881

Quoted 10 lines
kernel/audit.c +2 -5
 
-		read_lock(&tasklist_lock);
+		rcu_read_lock();
 		tsk = find_task_by_vpid(pid);
-		if (!tsk)
-			err = -ESRCH;
-		else {
-			spin_lock_irq(&tsk->sighand->siglock);
+		if (tsk && lock_task_sighand(tsk, &flags)) {
 			s.enabled = tsk->signal->audit_tty != 0;

Yes, this is what original code does, it takes ->siglock every time around read/write of ->audit_tty. And this looks absolutely bogus. Say, tty_audit_fork(). Why does it take ->siglock ?

Quoted 5 lines

As for ->tty_audit_buf, I am not sure ->siglock is the best choice, perhaps task_lock() would be better.

Once again, I think the patch is fine. Just it seems to me this code needs more cleanups.

Yeah, thought about that, but that's not in the scope of what I was working on. I leave that to the audit folks. :)

Thanks,

tglx

#6 lore raw

Re: [patch 3/3] audit: Use rcu for task lookup protection

Quoted 11 lines · Thomas Gleixner
Quoted 8 lines · Oleg Nesterov

But, sorry, can't resists ;) off-topic nit.

Quoted 6 lines
kernel/audit.c +1 -0
@@ -873,17 +873,16 @@ static int audit_receive_msg(struct sk_b
 	case AUDIT_TTY_GET: {
 		struct audit_tty_status s;
 		struct task_struct *tsk;
+		unsigned long flags;
 
Quoted 13 lines
Quoted 5 lines
Quoted 5 lines
kernel/audit.c +1 -3
-		read_lock(&tasklist_lock);
+		rcu_read_lock();
 		tsk = find_task_by_vpid(pid);
-		if (!tsk)
-			err = -ESRCH;
Quoted 8 lines

kernel/audit.c static int audit_receive_msg(struct sk_b · line 881

Quoted 4 lines
kernel/audit.c +1 -2
-		else {
-			spin_lock_irq(&tsk->sighand->siglock);
+		if (tsk && lock_task_sighand(tsk, &flags)) {
 			s.enabled = tsk->signal->audit_tty != 0;

Yes, this is what original code does, it takes ->siglock every time around read/write of ->audit_tty. And this looks absolutely bogus. Say, tty_audit_fork(). Why does it take ->siglock ?

Yes, I still think ->audit_tty doesn't need the locking.

Quoted 2 lines
Quoted 2 lines

As for ->tty_audit_buf, I am not sure ->siglock is the best choice, perhaps task_lock() would be better.

OOPS, I misread the code. ->tty_audit_buf is per-process (of course!). Well, unless I missed something again, tty_audit_push() and tty_audit_tiocsti() can access ->tty_audit_buf lockless.

Quoted 5 lines
Quoted 2 lines

Once again, I think the patch is fine. Just it seems to me this code needs more cleanups.

Yeah, thought about that, but that's not in the scope of what I was working on. I leave that to the audit folks. :)

Yes, yes, sure.

Oleg.