Re: [TOMOYO 15/15] LSM expansion for TOMOYO Linux.

Hello.

Thank you very much for your time, Paul.
Yes, you understood what I wanted to do.

TOMOYO Linux's approach:

(1) It uses userspace intervention to allow/reject
    connections and/or packets based on the application's domain.
    Since existent hooks can't be used for this purpose,
    I inserted a new hook post_recv_datagram() at skb_recv_datagram()
    and I modified socket_post_accept() to return error so that
    I can drop/disconnect based on the application's domain.

    I think skb_recv_datagram() is the only place that can remove
    a message picked up with MSG_PEEK flags from the receive queue.
    To remove a message picked up with MSG_PEEK flags, I noticed that
    I have to do skb_kill_datagram()-like operation so that
    "the head message that must not be delivered to the caller" won't prevent
    picking up of "the non-head message that should be delivered to the caller"
    when the caller repeats only recv(MSG_PEEK) requests.
    Since skb_recv_datagram() can be called from interrupt context,
    I have to use spin_lock_irqsave() instead for spin_lock_bh(), am I right?

/* from net/core/datagram.c */

@@ -178,6 +179,27 @@ struct sk_buff *skb_recv_datagram(struct
 		} else
 			skb = skb_dequeue(&sk->sk_receive_queue);
 
+		error = security_post_recv_datagram(sk, skb, flags);
+		if (error) {
+			unsigned long cpu_flags;
+
+			if (!(flags & MSG_PEEK))
+				goto no_peek;
+
+			spin_lock_irqsave(&sk->sk_receive_queue.lock,
+					  cpu_flags);
+			if (skb == skb_peek(&sk->sk_receive_queue)) {
+				__skb_unlink(skb,
+					     &sk->sk_receive_queue);
+				atomic_dec(&skb->users);
+			}
+			spin_unlock_irqrestore(&sk->sk_receive_queue.lock,
+					       cpu_flags);
+no_peek:
+			skb_free_datagram(sk, skb);
+			goto no_packet;
+		}
+
 		if (skb)
 			return skb;



    By the way, why can't socket_post_accept() fail?
    Someone may wish to do memory allocation at socket_post_accept().
    socket_accept() is too early for memory allocation because
    there is no chance to free allocated memory
    when sock->ops->accept() failed.
    I think socket_post_accept() should be able to fail.

(2) It allows the administrator judge interactively
    using a userspace agent.
    Thus, the new hook has to be inserted at blockable location,
    Since skb_recv_datagram() can be called from interrupt context,
    I do nothing in post_recv_datagram() if called from interrupt context.

+static int tmy_post_recv_datagram(struct sock *sk,
+				  struct sk_buff *skb,
+				  unsigned int flags)
+{
+	int error = 0;
+	const unsigned int type = sk->sk_type;
+
+	/* skb_recv_datagram() didn't dequeue. */
+	if (!skb)
+		return 0;
+
+	/* skb_recv_datagram() can be called from interrupt context. */
+	if (in_interrupt())
+		return 0;
+	/* I don't check if called by kernel process. */
+	if (segment_eq(get_fs(), KERNEL_DS))
+		return 0;
+
+	if (type != SOCK_DGRAM && type != SOCK_RAW)
+		return 0;
...(sniped)...
+}



Regards.

On Thursday, September 6 2007 9:04:01 am Tetsuo Handa wrote:

(1) It uses userspace intervention to allow/reject
    connections and/or packets based on the application's domain.
    Since existent hooks can't be used for this purpose,
    I inserted a new hook post_recv_datagram() at skb_recv_datagram()
    and I modified socket_post_accept() to return error so that
    I can drop/disconnect based on the application's domain.

    I think skb_recv_datagram() is the only place that can remove
    a message picked up with MSG_PEEK flags from the receive queue.
    To remove a message picked up with MSG_PEEK flags, I noticed that
    I have to do skb_kill_datagram()-like operation so that
    "the head message that must not be delivered to the caller" won't
prevent picking up of "the non-head message that should be delivered to the
caller" when the caller repeats only recv(MSG_PEEK) requests.
    Since skb_recv_datagram() can be called from interrupt context,
    I have to use spin_lock_irqsave() instead for spin_lock_bh(), am I
right?

There are almost certainly better people to answer locking questions, but here 
is my take on it ... If you are accessing data both in a bottom half and 
elsewhere you need to make sure you disable bottom halfs from running before 
you access the data outside the bottom half (spin_lock_bh()).  If you are 
accessing data both in an interrupt handler and elsewhere you need to make 
sure you disable interrupts when accessing data outside the irq handler 
(spin_lock_irqsave()).

    By the way, why can't socket_post_accept() fail?
    Someone may wish to do memory allocation at socket_post_accept().
    socket_accept() is too early for memory allocation because
    there is no chance to free allocated memory
    when sock->ops->accept() failed.
    I think socket_post_accept() should be able to fail.

From my experience the community disapproves of approaches which go through 

the entire TCP handshake and then terminate the connection, which is what 
allowing security_socket_post_accept() to fail would do.

-- 
paul moore
linux security @ hp