iwlwifi: don't count the tfds in HW queue any more

Since packets sent to an RA / TID in AGG are sent from a
separate HW Tx queue, we may get into a race:
the regular queue isn't empty while we already begin to
send packets from the AGG queue. This would result in sending
packets out of order.

In order to cope with this, mac80211 waits until the driver
reports that the legacy queue is drained before it can send
packets to the AGG queue. During that time, mac80211 buffers
packets for the driver. These packets will be sent in order
after the driver reports it is ready.

The way this was implemented in the driver is as follows:
We held a counter that monitors the number of packets for
an RA / TID in the HW queues. When this counter reached 0,
we knew that the HW queues were drained and we reported to
mac80211 that were ready to proceed.

This patch changes the implementation described above. We
now remember what is the wifi sequence number of the first
packet that will be sent in the AGG queue (lets' call it
ssn). When we reclaim the packet before ssn, we know that
the queue is drained, and we are ready to proceed.

This will allow us to move this logic in the upper layer and
eventually remove the tid_data from the shared area.

Signed-off-by: Emmanuel Grumbach <emmanuel.grumbach@intel.com>
Signed-off-by: Wey-Yi Guy <wey-yi.w.guy@intel.com>

1 files changed, 26 insertions, 2 deletions

diff --git a/drivers/net/wireless/iwlwifi/iwl-agn-tx.c b/drivers/net/wireless/iwlwifi/iwl-agn-tx.c
index 1cac701c418..69d0f997298 100644
--- a/drivers/net/wireless/iwlwifi/iwl-agn-tx.c
+++ b/drivers/net/wireless/iwlwifi/iwl-agn-tx.c
@@ -772,7 +772,7 @@ int iwlagn_rx_reply_tx(struct iwl_priv *priv, struct iwl_rx_mem_buffer *rxb,
 	struct iwlagn_tx_resp *tx_resp = (void *)&pkt->u.raw[0];
 	struct ieee80211_hdr *hdr;
 	u32 status = le16_to_cpu(tx_resp->status.status);
-	u32 ssn = iwlagn_get_scd_ssn(tx_resp);
+	u16 ssn = iwlagn_get_scd_ssn(tx_resp);
 	int tid;
 	int sta_id;
 	int freed;
@@ -794,8 +794,31 @@ int iwlagn_rx_reply_tx(struct iwl_priv *priv, struct iwl_rx_mem_buffer *rxb,
 		iwl_rx_reply_tx_agg(priv, tx_resp);
 
 	if (tx_resp->frame_count == 1) {
-		IWL_DEBUG_TX_REPLY(priv, "Q %d, ssn %d", txq_id, ssn);
+		u16 next_reclaimed = le16_to_cpu(tx_resp->seq_ctl);
+		next_reclaimed = SEQ_TO_SN(next_reclaimed + 0x10);
+
+		if (is_agg) {
+			/* If this is an aggregation queue, we can rely on the
+			 * ssn since the wifi sequence number corresponds to
+			 * the index in the TFD ring (%256).
+			 * The seq_ctl is the sequence control of the packet
+			 * to which this Tx response relates. But if there is a
+			 * hole in the bitmap of the BA we received, this Tx
+			 * response may allow to reclaim the hole and all the
+			 * subsequent packets that were already acked.
+			 * In that case, seq_ctl != ssn, and the next packet
+			 * to be reclaimed will be ssn and not seq_ctl.
+			 */
+			next_reclaimed = ssn;
+		}
+
 		__skb_queue_head_init(&skbs);
+		priv->shrd->tid_data[sta_id][tid].next_reclaimed =
+			next_reclaimed;
+
+		IWL_DEBUG_TX_REPLY(priv, "Next reclaimed packet:%d",
+					  next_reclaimed);
+
 		/*we can free until ssn % q.n_bd not inclusive */
 		iwl_trans_reclaim(trans(priv), sta_id, tid, txq_id,
 				  ssn, status, &skbs);
@@ -951,6 +974,7 @@ int iwlagn_rx_reply_compressed_ba(struct iwl_priv *priv,
 	/* Release all TFDs before the SSN, i.e. all TFDs in front of
 	 * block-ack window (we assume that they've been successfully
 	 * transmitted ... if not, it's too late anyway). */
+	priv->shrd->tid_data[sta_id][tid].next_reclaimed = ba_resp_scd_ssn;
 	iwl_trans_reclaim(trans(priv), sta_id, tid, scd_flow, ba_resp_scd_ssn,
 			  0, &reclaimed_skbs);
 	freed = 0;