mt76: mt7663s: introduce sdio tx aggregation

Introduce sdio tx aggregation to reduce bus transaction ands improve tx
throughput. For the moment the skb are copied in a dedicated buffer
since mmc APIs do not support sg table for zero-copy.
Since skb data are already copied in xmit_buff[], avoid linearization in
ma80211 layer. Relying on tx aggregation, we improve tx tpt of ~65%.

Tested-by: Sean Wang <sean.wang@mediatek.com>
Co-developed-by: Sean Wang <sean.wang@mediatek.com>
Signed-off-by: Sean Wang <sean.wang@mediatek.com>
Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>
Signed-off-by: Felix Fietkau <nbd@nbd.name>

2 files changed, 46 insertions, 15 deletions

diff --git a/drivers/net/wireless/mediatek/mt76/mt7615/sdio.c b/drivers/net/wireless/mediatek/mt76/mt7615/sdio.c
index 8621c6f579aa..874c929d8552 100644
--- a/drivers/net/wireless/mediatek/mt76/mt7615/sdio.c
+++ b/drivers/net/wireless/mediatek/mt76/mt7615/sdio.c
@@ -346,7 +346,7 @@ static int mt7663s_probe(struct sdio_func *func,
 	struct ieee80211_ops *ops;
 	struct mt7615_dev *dev;
 	struct mt76_dev *mdev;
-	int ret;
+	int i, ret;
 
 	ops = devm_kmemdup(&func->dev, &mt7615_ops, sizeof(mt7615_ops),
 			   GFP_KERNEL);
@@ -387,6 +387,16 @@ static int mt7663s_probe(struct sdio_func *func,
 		goto err_deinit;
 	}
 
+	for (i = 0; i < ARRAY_SIZE(mdev->sdio.xmit_buf); i++) {
+		mdev->sdio.xmit_buf[i] = devm_kmalloc(mdev->dev,
+						      MT76S_XMIT_BUF_SZ,
+						      GFP_KERNEL);
+		if (!mdev->sdio.xmit_buf[i]) {
+			ret = -ENOMEM;
+			goto err_deinit;
+		}
+	}
+
 	ret = mt76s_alloc_queues(&dev->mt76);
 	if (ret)
 		goto err_deinit;
diff --git a/drivers/net/wireless/mediatek/mt76/mt7615/sdio_txrx.c b/drivers/net/wireless/mediatek/mt76/mt7615/sdio_txrx.c
index 4033fe431312..e82f6bdcbce4 100644
--- a/drivers/net/wireless/mediatek/mt76/mt7615/sdio_txrx.c
+++ b/drivers/net/wireless/mediatek/mt76/mt7615/sdio_txrx.c
@@ -138,15 +138,11 @@ static int mt7663s_rx_run_queue(struct mt76_dev *dev, enum mt76_rxq_id qid,
 	return i;
 }
 
-static int mt7663s_tx_pick_quota(struct mt76_dev *dev, enum mt76_txq_id qid,
+static int mt7663s_tx_pick_quota(struct mt76_sdio *sdio, enum mt76_txq_id qid,
 				 int buf_sz, int *pse_size, int *ple_size)
 {
-	struct mt76_sdio *sdio = &dev->sdio;
 	int pse_sz;
 
-	if (!test_bit(MT76_STATE_MCU_RUNNING, &dev->phy.state))
-		return 0;
-
 	pse_sz = DIV_ROUND_UP(buf_sz + sdio->sched.deficit, MT_PSE_PAGE_SZ);
 
 	if (qid == MT_TXQ_MCU) {
@@ -197,27 +193,52 @@ static int __mt7663s_xmit_queue(struct mt76_dev *dev, u8 *data, int len)
 
 static int mt7663s_tx_run_queue(struct mt76_dev *dev, enum mt76_txq_id qid)
 {
-	int nframes = 0, pse_sz = 0, ple_sz = 0;
+	int err, nframes = 0, len = 0, pse_sz = 0, ple_sz = 0;
 	struct mt76_queue *q = dev->q_tx[qid];
 	struct mt76_sdio *sdio = &dev->sdio;
 
 	while (q->first != q->head) {
 		struct mt76_queue_entry *e = &q->entry[q->first];
-		int err;
+		struct sk_buff *iter;
+
+		if (!test_bit(MT76_STATE_MCU_RUNNING, &dev->phy.state)) {
+			__skb_put_zero(e->skb, 4);
+			err = __mt7663s_xmit_queue(dev, e->skb->data,
+						   e->skb->len);
+			if (err)
+				return err;
+
+			goto next;
+		}
+
+		if (len + e->skb->len + 4 > MT76S_XMIT_BUF_SZ)
+			break;
 
-		if (mt7663s_tx_pick_quota(dev, qid, e->buf_sz, &pse_sz,
+		if (mt7663s_tx_pick_quota(sdio, qid, e->buf_sz, &pse_sz,
 					  &ple_sz))
 			break;
 
-		__skb_put_zero(e->skb, 4);
+		memcpy(sdio->xmit_buf[qid] + len, e->skb->data,
+		       skb_headlen(e->skb));
+		len += skb_headlen(e->skb);
+		nframes++;
 
-		err = __mt7663s_xmit_queue(dev, e->skb->data, e->skb->len);
+		skb_walk_frags(e->skb, iter) {
+			memcpy(sdio->xmit_buf[qid] + len, iter->data,
+			       iter->len);
+			len += iter->len;
+			nframes++;
+		}
+next:
+		q->first = (q->first + 1) % q->ndesc;
+		e->done = true;
+	}
+
+	if (nframes) {
+		memset(sdio->xmit_buf[qid] + len, 0, 4);
+		err = __mt7663s_xmit_queue(dev, sdio->xmit_buf[qid], len + 4);
 		if (err)
 			return err;
-
-		e->done = true;
-		q->first = (q->first + 1) % q->ndesc;
-		nframes++;
 	}
 	mt7663s_tx_update_quota(sdio, qid, pse_sz, ple_sz);