Commit 3995bd93 authored by Johannes Berg's avatar Johannes Berg Committed by John W. Linville
Browse files

iwlwifi: fix TX queue race

I had a problem on 4965 hardware (well, probably other hardware too,
but others don't survive my stress testing right now, unfortunately)
where the driver was sending invalid commands to the device, but no
such thing could be seen from the driver's point of view. I could
reproduce this fairly easily by sending multiple TCP streams with
iperf on different TIDs, though sometimes a single iperf stream was
sufficient. It even happened with a single core, but I have forced
preemption turned on.

The culprit was a queue overrun, where we advanced the queue's write
pointer over the read pointer. After careful analysis I've come to
the conclusion that the cause is a race condition between iwlwifi
and mac80211.

mac80211, of course, checks whether the queue is stopped, before
transmitting a frame. This effectively looks like this:

        if (stopped(queue)) {
                return busy;
        ...             <-- this place will be important
			    there is some more code here

The driver, on the other hand, can stop and start queues, which does


	[if marked running: wake up tasklet to send pending frames]

Now, however, once the driver starts the queue, mac80211 can see that
and end up at the marked place above, at which point for some reason the
driver seems to stop the queue again (I don't understand that) and then
we end up transmitting while the queue is actually full.

Now, this shouldn't actually matter much, but for some reason I've seen
it happen multiple times in a row and the queue actually overflows, at
which point the queue bites itself in the tail and things go completely

This patch fixes this by just dropping the packet should this have
happened, and making the lock in iwlwifi cover everything so iwlwifi
can't race against itself (dropping the lock there might make it more
likely, but it did seem to happen without that too).

Since we can't hold the lock across drv_tx() above, I see no way to fix
this in mac80211, but I also don't understand why I haven't seen this
before -- maybe I just never stress tested it this badly.

With this patch, the device has survived many minutes of simultanously
sending two iperf streams on different TIDs with combined throughput
of about 60 Mbps.
Signed-off-by: default avatarJohannes Berg <>
Signed-off-by: default avatarReinette Chatre <>
Signed-off-by: default avatarJohn W. Linville <>
parent 8bae1b2b
......@@ -720,8 +720,6 @@ int iwl_tx_skb(struct iwl_priv *priv, struct sk_buff *skb)
goto drop_unlock;
spin_unlock_irqrestore(&priv->lock, flags);
hdr_len = ieee80211_hdrlen(fc);
/* Find (or create) index into station table for destination station */
......@@ -729,7 +727,7 @@ int iwl_tx_skb(struct iwl_priv *priv, struct sk_buff *skb)
if (sta_id == IWL_INVALID_STATION) {
IWL_DEBUG_DROP(priv, "Dropping - INVALID STATION: %pM\n",
goto drop;
goto drop_unlock;
IWL_DEBUG_TX(priv, "station Id %d\n", sta_id);
......@@ -750,14 +748,17 @@ int iwl_tx_skb(struct iwl_priv *priv, struct sk_buff *skb)
txq_id = priv->stations[sta_id].tid[tid].agg.txq_id;
swq_id = iwl_virtual_agg_queue_num(swq_id, txq_id);
txq = &priv->txq[txq_id];
q = &txq->q;
txq->swq_id = swq_id;
spin_lock_irqsave(&priv->lock, flags);
if (unlikely(iwl_queue_space(q) < q->high_mark))
goto drop_unlock;
if (ieee80211_is_data_qos(fc))
/* Set up driver data for this TFD */
memset(&(txq->txb[q->write_ptr]), 0, sizeof(struct iwl_tx_info));
......@@ -902,7 +903,6 @@ int iwl_tx_skb(struct iwl_priv *priv, struct sk_buff *skb)
spin_unlock_irqrestore(&priv->lock, flags);
return -1;
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment