On Wed, Dec 26, 2012 at 06:08:50 +0800, Marek Lindner wrote:
On Saturday, December 15, 2012 19:14:42 Antonio Quartulli wrote:
OGMs are currently prepared 1 originator interval in advance then the time they are used to be sent. This means that once in the air they carry old information (like TT announcements and possibly other flags). To fix this, postpone the OGM creation to the same time of sending, in this way the OGM is first created and then immediately sent.
How about: OGMs are currently assembled and enqueued one originator interval before they are sent. This means they might carry information which was outdated while the OGM waited in the outgoing packet queue (like TT announcements and possibly other flags). The OGM assembly has to be postponed to the latest possible moment before sending in order to minimize the gap between gathering the data and flooding it to the network.
Sounds much better. Thanks!
-/* when do we schedule our own ogm to be sent */ +/**
- batadv_iv_ogm_emit_wait_time - compute the OGM preparation waiting time
- @bat_priv: the bat priv with all the soft interface information
- Returns the amount of jiffies to wait before preparing and sending the
next + * own OGM
- */
I'd say "Returns the number of jiffies [..]".
Yeah
@@ -468,6 +442,7 @@ static void batadv_iv_ogm_aggregate(struct batadv_forw_packet *forw_packet_aggr, memcpy(skb_buff, packet_buff, packet_len); forw_packet_aggr->packet_len += packet_len; forw_packet_aggr->num_packets++;
forw_packet_aggr->own |= own_packet;
/* save packet direct link flag status */ if (direct_link) {
Using "|= own_packet" isn't strictly necessary because "forw_packet_aggr->own" isn't a bit field.
Well the point is that here own_packet could be false, but forw_packet_aggr->own might already be true, so I didn't want to destroy the original value.
Did you vigorously test this code ? Especially, multi-node with multiple interface setups are of interest. Also use different orig intervals to ensure it still works everywhere.
I will try more topologies and in particular different orig intervals as soon as I have the possibility
The thing is: Throughout the code you can find the implicite assumption of the first aggregated packet being an "own packet" (if forw_packet_aggr->own is set). Therefore, you have to be very careful changing that logic. One function you definitely overlooked is batadv_iv_ogm_send_to_if() but there might be others.
@@ -498,8 +473,7 @@ static void batadv_iv_ogm_queue_add(struct batadv_priv *bat_priv,
/* find position for the packet in the forward queue */ spin_lock_bh(&bat_priv->forw_bat_list_lock);
- /* own packets are not to be aggregated */
- if ((atomic_read(&bat_priv->aggregated_ogms)) && (!own_packet)) {
- if ((atomic_read(&bat_priv->aggregated_ogms))) { hlist_for_each_entry(forw_packet_pos, tmp_node, &bat_priv->forw_bat_list, list) { if (batadv_iv_ogm_can_aggregate(batadv_ogm_packet,
Now it gets interesting! Did you review the impact of batadv_iv_ogm_can_aggregate() before you bravely removed the exclusion of own packets ? That is quite a beast ...
well, the poor existing doc didn't point out all the possible threats :-P I'll investigate more. Thanks :)
+/**
- batadv_iv_ogm_send - prepare an send an own OGM
- @work: kernel work struct
- Prepare the OGM and immediately enqueue it for sending
- */
prepare and send an own OGM ^^^
well, technically it is enqueued, not sent..
+static void batadv_iv_ogm_send(struct work_struct *work)
The naming is less than optimal. bat_iv_ogm.c now has:
- batadv_iv_ogm_send
- batadv_iv_ogm_emit
mh..ok, I will propose something better :)
--- a/types.h +++ b/types.h @@ -49,6 +49,7 @@ struct batadv_hard_iface_bat_iv { unsigned char *ogm_buff; int ogm_buff_len; atomic_t ogm_seqno;
- struct delayed_work work;
};
Guess we need some kernel doc for this change too. Since types.h now is fully documented ...
only because you are commenting this patch after having merged your full-kernel-doc-patch :-D Ok, I will fix it too.
Thank you very much for your precious feedback. Greetings from down under!
Cheers,