Re: [RFC 2/2] Bluetooth: L2CAP ERTM state machine replacement

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On 2/24/2012 12:13 PM, Ulisses Furquim wrote:
Hi Mat,

On Thu, Feb 23, 2012 at 6:37 PM, Mat Martineau<mathewm@xxxxxxxxxxxxxx>  wrote:
The previous ERTM implementation had a handler for each frame type,
and each handler had to figure out what the current state was and
handle each frame accordingly.

This implementation has a state machine that chooses an execution path
first based on the receive or transmit state, then each state has
handlers for the frame types. This is easier to match up with the
spec, which is defined similarly.

Signed-off-by: Mat Martineau<mathewm@xxxxxxxxxxxxxx>

<snip>

@@ -1245,7 +1457,8 @@ int __l2cap_wait_ack(struct sock *sk)

         add_wait_queue(sk_sleep(sk),&wait);
         set_current_state(TASK_INTERRUPTIBLE);
-       while (chan->unacked_frames>  0&&  chan->conn) {
+       while (chan->unacked_frames>  0&&  chan->conn&&
+               atomic_read(&chan->ertm_queued)) {
                 if (!timeo)
                         timeo = HZ/5;

Can we have unacked_frames>  0 and nothing queued? Or have I misinterpreted?

In normal operation, there should be unacked frames when ertm_queued is non-zero. I think I ran in to a corner case with AMP, where unacked_frames can be forced to 0 during a channel move. There are AMP components to the state machine that are not included in this patch.


<snip>

+               BT_DBG("Sent txseq %d", (int)control->txseq);

-               skb = skb_queue_next(&chan->tx_q, skb);
+               chan->next_tx_seq = __next_seq(chan->next_tx_seq, chan);
+               chan->frames_sent += 1;
+               sent += 1;

Nitpick here. frames_sent++ and sent++ ? Happens in other places as
well so I won't copy all of them here.

Ok, will fix.


<snip>

-               if (bt_cb(skb)->retries == 1) {
-                       chan->unacked_frames++;
+               l2cap_chan_hold(chan);
+               sock_hold(chan->sk);
+               tx_skb->sk = chan->sk;

Do we really need these? Do we always have chan->sk? (We have that in
l2cap_ertm_send() and l2cap_ertm_resend()).

The upstream ERTM code still relies on having chan->sk, so I didn't try to finish splitting channels & sockets within this patch. skb destructors expect to have an sk pointer, so it is critical to modify these reference counts so the socket and chan are around when the skb leaves the hci tx queue.


<snip>

+               keep_sk = schedule_work(&chan->tx_work);

Would make sense to schedule this in hdev->workqueue?

It's a tradeoff. If this is scheduled on hdev->workqueue, then that workqueue can get blocked waiting for the socket lock. If these are scheduled on the system workqueue, there is potential for more latency (but it hasn't been a problem in practice, even with AMP data rates).


<snip>

+static void l2cap_ertm_tx_worker(struct work_struct *work)
  {

Why do we need this worker?

To control the depth of the hci tx queue. Without this, you end up with an entire tx window of iframes queued up at the hci layer. When an sframe shows up from the remote device and you have to retransmit, or when an sframe needs to be sent, then retransmissions and sframes have to get queued behind that full tx window of iframes. It adds a HUGE amount of latency in those situations, which leads to ERTM timeouts and disconnects that are not necessary. ERTM works much, much better with lossy connections (like AMP) if it does not flood the hci tx queue.

We had a discussion on the list about how to solve this. The idea is to push most queuing up to the L2CAP layer, and have the hci scheduler call up to L2CAP to fetch frames. However, this hasn't been implemented yet.


-       int ret;
+       struct l2cap_chan *chan =
+               container_of(work, struct l2cap_chan, tx_work);

-       if (!skb_queue_empty(&chan->tx_q))
-               chan->tx_send_head = chan->tx_q.next;
+       BT_DBG("%p", chan);

-       chan->next_tx_seq = chan->expected_ack_seq;
-       ret = l2cap_ertm_send(chan);
-       return ret;
+       lock_sock(chan->sk);
+       l2cap_ertm_send(chan);
+       release_sock(chan->sk);

Can't we just use l2cap_chan_lock()/l2cap_chan_unlock() here? (I'm
assuming you saw the patches creating a lock in l2cap_chan to protect
it) Do we always have sk?

l2cap_chan_lock() is pretty new! Yes, I should use that to guard the ERTM state.

Right now, we do still always have sk, but that is changing (as it should).

+       sock_put(chan->sk);
+       l2cap_chan_put(chan);
  }

<snip>

+static void l2cap_monitor_timeout(struct work_struct *work)
+{
+       struct l2cap_chan *chan = container_of(work, struct l2cap_chan,
+                                                       monitor_timer.work);
+       struct sock *sk = chan->sk;
+
+       BT_DBG("chan %p", chan);
+
+       lock_sock(sk);
+
+       if (!chan->conn) {
+               release_sock(sk);
+               l2cap_chan_put(chan);
+               return;
+       }
+
+       l2cap_ertm_tx(chan, 0, 0, L2CAP_ERTM_EVENT_MONITOR_TIMER_EXPIRES);
+
+       release_sock(sk);
+       l2cap_chan_put(chan);
+}

Here we might need to use l2cap_chan_lock/unlock instead of
lock_sock/release_sock.

Agreed.


+static void l2cap_retrans_timeout(struct work_struct *work)
+{
+       struct l2cap_chan *chan = container_of(work, struct l2cap_chan,
+                                                       retrans_timer.work);
+       struct sock *sk = chan->sk;
+
+       BT_DBG("chan %p", chan);
+
+       lock_sock(sk);
+
+       if (!chan->conn) {
+               release_sock(sk);
+               l2cap_chan_put(chan);
+               return;
+       }
+
+       l2cap_ertm_tx(chan, 0, 0, L2CAP_ERTM_EVENT_RETRANS_TIMER_EXPIRES);
+       release_sock(sk);
+       l2cap_chan_put(chan);

And here as well.

Ok, that's it for now. Have you run this code through PTS or any other
test? I haven't checked the actual bits of ERTM but since we're
replacing the current state machine code we need to be somewhat sure
this code is as good as or even better than the current one.
Introducing too many regressions would be really bad IMHO.

The kernel I'm porting from is qualified, listed, interop'd, UPF'd, and heavily tested -- but I haven't validated this port yet. I will definitely run this through PTS and test against other ERTM implementations before we merge the changes.

--
Mat Martineau
Employee of Qualcomm Innovation Center, Inc.
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum
--
To unsubscribe from this list: send the line "unsubscribe linux-bluetooth" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html


[Index of Archives]     [Bluez Devel]     [Linux Wireless Networking]     [Linux Wireless Personal Area Networking]     [Linux ATH6KL]     [Linux USB Devel]     [Linux Media Drivers]     [Linux Audio Users]     [Linux Kernel]     [Linux SCSI]     [Big List of Linux Books]

  Powered by Linux