ath10k: implement push-pull tx

The current/old tx path design was that host, at
its own leisure, pushed tx frames to the device.
For HTT there was ~1000-1400 msdu queue depth.

After reaching that limit the driver would request
mac80211 to stop queues. There was little control
over what packets got in there as far as
DA/RA was considered so it was rather easy to
starve per-station traffic flows.

With MU-MIMO this became a significant problem
because the queue depth was insufficient to buffer
frames from multiple clients (which could have
different signal quality and capabilities) in an
efficient fashion.

Hence the new tx path in 10.4 was introduced: a
pull-push mode.

Firmware and host can share tx queue state via
DMA. The state is logically a 2 dimensional array
addressed via peer_id+tid pair. Each entry is a
counter (either number of bytes or packets. Host
keeps it updated and firmware uses it for
scheduling Tx pull requests to host.

This allows MU-MIMO to become a lot more effective
with 10+ clients.

Signed-off-by: Michal Kazior <michal.kazior@tieto.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>

1 files changed, 29 insertions, 10 deletions

diff --git a/drivers/net/wireless/ath/ath10k/htt_tx.c b/drivers/net/wireless/ath/ath10k/htt_tx.c
index 6643be8692b5..a30c34eae0a7 100644
--- a/drivers/net/wireless/ath/ath10k/htt_tx.c
+++ b/drivers/net/wireless/ath/ath10k/htt_tx.c
@@ -64,6 +64,9 @@ static void __ath10k_htt_tx_txq_recalc(struct ieee80211_hw *hw,
 	if (!ar->htt.tx_q_state.enabled)
 		return;
 
+	if (ar->htt.tx_q_state.mode != HTT_TX_MODE_SWITCH_PUSH_PULL)
+		return;
+
 	if (txq->sta)
 		peer_id = arsta->peer_id;
 	else
@@ -101,6 +104,9 @@ static void __ath10k_htt_tx_txq_sync(struct ath10k *ar)
 	if (!ar->htt.tx_q_state.enabled)
 		return;
 
+	if (ar->htt.tx_q_state.mode != HTT_TX_MODE_SWITCH_PUSH_PULL)
+		return;
+
 	seq = le32_to_cpu(ar->htt.tx_q_state.vaddr->seq);
 	seq++;
 	ar->htt.tx_q_state.vaddr->seq = cpu_to_le32(seq);
@@ -115,6 +121,23 @@ static void __ath10k_htt_tx_txq_sync(struct ath10k *ar)
 				   DMA_TO_DEVICE);
 }
 
+void ath10k_htt_tx_txq_recalc(struct ieee80211_hw *hw,
+			      struct ieee80211_txq *txq)
+{
+	struct ath10k *ar = hw->priv;
+
+	spin_lock_bh(&ar->htt.tx_lock);
+	__ath10k_htt_tx_txq_recalc(hw, txq);
+	spin_unlock_bh(&ar->htt.tx_lock);
+}
+
+void ath10k_htt_tx_txq_sync(struct ath10k *ar)
+{
+	spin_lock_bh(&ar->htt.tx_lock);
+	__ath10k_htt_tx_txq_sync(ar);
+	spin_unlock_bh(&ar->htt.tx_lock);
+}
+
 void ath10k_htt_tx_txq_update(struct ieee80211_hw *hw,
 			      struct ieee80211_txq *txq)
 {
@@ -638,10 +661,14 @@ int ath10k_htt_tx_fetch_resp(struct ath10k *ar,
 {
 	struct sk_buff *skb;
 	struct htt_cmd *cmd;
-	u16 resp_id;
+	const u16 resp_id = 0;
 	int len = 0;
 	int ret;
 
+	/* Response IDs are echo-ed back only for host driver convienence
+	 * purposes. They aren't used for anything in the driver yet so use 0.
+	 */
+
 	len += sizeof(cmd->hdr);
 	len += sizeof(cmd->tx_fetch_resp);
 	len += sizeof(cmd->tx_fetch_resp.records[0]) * num_records;
@@ -650,11 +677,6 @@ int ath10k_htt_tx_fetch_resp(struct ath10k *ar,
 	if (!skb)
 		return -ENOMEM;
 
-	resp_id = 0; /* TODO: allocate resp_id */
-	ret = 0;
-	if (ret)
-		goto err_free_skb;
-
 	skb_put(skb, len);
 	cmd = (struct htt_cmd *)skb->data;
 	cmd->hdr.msg_type = HTT_H2T_MSG_TYPE_TX_FETCH_RESP;
@@ -669,14 +691,11 @@ int ath10k_htt_tx_fetch_resp(struct ath10k *ar,
 	ret = ath10k_htc_send(&ar->htc, ar->htt.eid, skb);
 	if (ret) {
 		ath10k_warn(ar, "failed to submit htc command: %d\n", ret);
-		goto err_free_resp_id;
+		goto err_free_skb;
 	}
 
 	return 0;
 
-err_free_resp_id:
-	(void)resp_id; /* TODO: free resp_id */
-
 err_free_skb:
 	dev_kfree_skb_any(skb);