On Tue, 9 Aug 2011, Gustavo Padovan wrote:
* Mat Martineau <mathewm@xxxxxxxxxxxxxx> [2011-08-08 16:29:51 -0700]:
On Fri, 5 Aug 2011, Gustavo Padovan wrote:
* Peter Hurley <peter@xxxxxxxxxxxxxxxxxx> [2011-08-04 19:09:37 -0400]:
Hi Mat,
On Thu, 2011-08-04 at 13:37 -0400, Mat Martineau wrote:
I had a recent discussion with Gustavo about HCI queuing issues with
ERTM:
http://www.spinics.net/lists/linux-bluetooth/msg13774.html
My proposal is to move tx queuing up to L2CAP, and have the HCI tx
task only handle scheduling. Senders would tell HCI they have data to
send, and HCI would call back to pull data. I've been focused on
L2CAP - it would be possible to make a similar queuing change to
SCO/eSCO/LE, but not strictly necessary.
Would you please clarify this approach (perhaps in a separate thread)?
For example, how does having tx queues in l2cap_chan (instead of the
hci_conn) solve the latency problems in ERTM when replying to
REJ/SREJ/poll? Won't there potentially be just as much data already
queued up? Is the plan to move the reply to the front of the tx queue
because reqseq won't need to be assigned until the frame is actually
pulled off the queue?
Exactly. ERTM connections can get dropped if the too much data is buffered and
we need to send final bit for example.
Right now, an outgoing ERTM frame goes through two queues: a
channel-specific ERTM tx queue and the HCI ACL data_q. The ERTM
control field is not constructed until a frame is removed from the
ERTM tx queue and pushed to the HCI data_q, so the s-frame latency
problem comes in when the the HCI data_q gets deep. S-frames are
already pushed directly in to the HCI data_q, bypassing the data tx
queue.
From an ERTM perspective, the goal is to defer assignment of reqseq
and f-bit values as late as possible, so the remote device gets the
most recent information on data frames and polls that have been
received. The optimal thing to do (by this measurement, anyway) is
to build the ERTM control field as data is sent to the baseband --
in other words, to eliminate the HCI data_q altogether.
(Yeah, without the data_q, ERTM would need additional queues for
s-frames and retransmitted i-frames)
So, without a data_q, what makes sense? If there are ACL buffers
available and no pending L2CAP senders, it would be great to push
data straight out to the baseband. If we're blocked waiting for
num_completed_packets, then receipt of num_completed_packets is the
natural time to pull data from the tx queues that now happen to be
up in the L2CAP layer.
There are certainly locking, task scheduling, data scheduling, QoS,
and efficiency issues to consider. This is just a general
description for now, and I'm trying to see if there's enough
interest (or few enough obvious gotchas) to put some serious effort
in to moving forward.
Getting rid of conn->data_q makes sense. I started a patch to create the
struct hci_chan that Luiz proposed. It would be one HCI channel per L2CAP
connection. The buffer (acl_cnt) would be now divided by the number of
channels and not the number of connections. This is a first step to support
QoS and priority inside ERTM. QoS then would just need new scheduler rules.
+struct hci_chan {
+ struct list_head list;
+ struct hci_conn *conn;
+ struct sk_buff_head data_q;
+ unsigned int sent;
+}
At the BlueZ summit last year, the group settled on using an "HCI
channel" as an abstraction for AMP logical links. We have a hci_chan
struct already that you could add to. Before making changes to HCI
data structures, could we first work on upstreaming the AMP HCI
changes (which also include some QoS-related code)?
So in the next step for ERTM we move the queue to L2CAP and create a callback
to call from HCI at the moment of push data to the baseband. The function in
L2CAP would set the last control bits in the first packet of the queue and
sent it through.
This actually causes some problems for ERTM, since skbs are cloned
before they are pushed to HCI. skb data is not supposed to be
modified after cloning.
If there's a callback to L2CAP anyway, why not have L2CAP provide the
skb at that time instead of modifying data it provided earlier?
Then the queue can be split in two by adding a pointer that will mark which
element divides the queue between prio and normal. New prio skbs would just be
queued after this element and before the rest.
I think it's simpler and less bug-prone to just have two queues.
Either way, it's one more pointer.
However, I'm still not sure we want any queues in hci_chan. It's not
very complicated to have the queue in the L2CAP layer, and gives ERTM
the control it needs.
I still need to think on locking here. (and also finish my patches
that move all the bluetooth to workqueue)
Keep in mind that skb_queue() and skb_dequeue() have their own
locking. Non-ERTM modes shouldn't need locking when HCI calls back for
skbs.
For ERTM, we need to figure out a good way to protect ERTM state (like
buffer_seq and next_tx_seq) without using the socket lock.
--
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
