mt76: add support for overriding the device used for DMA mapping

WED support requires using non-coherent DMA, whereas the PCI device might
be configured for coherent DMA.
The WED driver will take care of changing the PCI HIF coherent IO setting
on attach.
Signed-off-by: Felix Fietkau <nbd@nbd.name>

drivers/net/wireless/mediatek/mt76/dma.c

@@ -24,7 +24,7 @@ mt76_alloc_txwi(struct mt76_dev *dev)
 	if (!txwi)
 		return NULL;
 
-	addr = dma_map_single(dev->dev, txwi, dev->drv->txwi_size,
+	addr = dma_map_single(dev->dma_dev, txwi, dev->drv->txwi_size,
 			      DMA_TO_DEVICE);
 	t = (struct mt76_txwi_cache *)(txwi + dev->drv->txwi_size);
 	t->dma_addr = addr;
@@ -78,7 +78,7 @@ mt76_free_pending_txwi(struct mt76_dev *dev)
 
 	local_bh_disable();
 	while ((t = __mt76_get_txwi(dev)) != NULL) {
-		dma_unmap_single(dev->dev, t->dma_addr, dev->drv->txwi_size,
+		dma_unmap_single(dev->dma_dev, t->dma_addr, dev->drv->txwi_size,
 				 DMA_TO_DEVICE);
 		kfree(mt76_get_txwi_ptr(dev, t));
 	}
@@ -127,7 +127,7 @@ mt76_dma_alloc_queue(struct mt76_dev *dev, struct mt76_queue *q,
 	q->hw_idx = idx;
 
 	size = q->ndesc * sizeof(struct mt76_desc);
-	q->desc = dmam_alloc_coherent(dev->dev, size, &q->desc_dma, GFP_KERNEL);
+	q->desc = dmam_alloc_coherent(dev->dma_dev, size, &q->desc_dma, GFP_KERNEL);
 	if (!q->desc)
 		return -ENOMEM;
 
@@ -209,11 +209,11 @@ mt76_dma_tx_cleanup_idx(struct mt76_dev *dev, struct mt76_queue *q, int idx,
 	struct mt76_queue_entry *e = &q->entry[idx];
 
 	if (!e->skip_buf0)
-		dma_unmap_single(dev->dev, e->dma_addr[0], e->dma_len[0],
+		dma_unmap_single(dev->dma_dev, e->dma_addr[0], e->dma_len[0],
 				 DMA_TO_DEVICE);
 
 	if (!e->skip_buf1)
-		dma_unmap_single(dev->dev, e->dma_addr[1], e->dma_len[1],
+		dma_unmap_single(dev->dma_dev, e->dma_addr[1], e->dma_len[1],
 				 DMA_TO_DEVICE);
 
 	if (e->txwi == DMA_DUMMY_DATA)
@@ -293,7 +293,7 @@ mt76_dma_get_buf(struct mt76_dev *dev, struct mt76_queue *q, int idx,
 	if (info)
 		*info = le32_to_cpu(desc->info);
 
-	dma_unmap_single(dev->dev, buf_addr, buf_len, DMA_FROM_DEVICE);
+	dma_unmap_single(dev->dma_dev, buf_addr, buf_len, DMA_FROM_DEVICE);
 	e->buf = NULL;
 
 	return buf;
@@ -330,9 +330,9 @@ mt76_dma_tx_queue_skb_raw(struct mt76_dev *dev, struct mt76_queue *q,
 	if (q->queued + 1 >= q->ndesc - 1)
 		goto error;
 
-	addr = dma_map_single(dev->dev, skb->data, skb->len,
+	addr = dma_map_single(dev->dma_dev, skb->data, skb->len,
 			      DMA_TO_DEVICE);
-	if (unlikely(dma_mapping_error(dev->dev, addr)))
+	if (unlikely(dma_mapping_error(dev->dma_dev, addr)))
 		goto error;
 
 	buf.addr = addr;
@@ -379,8 +379,8 @@ mt76_dma_tx_queue_skb(struct mt76_dev *dev, struct mt76_queue *q,
 		mt76_insert_hdr_pad(skb);
 
 	len = skb_headlen(skb);
-	addr = dma_map_single(dev->dev, skb->data, len, DMA_TO_DEVICE);
-	if (unlikely(dma_mapping_error(dev->dev, addr)))
+	addr = dma_map_single(dev->dma_dev, skb->data, len, DMA_TO_DEVICE);
+	if (unlikely(dma_mapping_error(dev->dma_dev, addr)))
 		goto free;
 
 	tx_info.buf[n].addr = t->dma_addr;
@@ -392,9 +392,9 @@ mt76_dma_tx_queue_skb(struct mt76_dev *dev, struct mt76_queue *q,
 		if (n == ARRAY_SIZE(tx_info.buf))
 			goto unmap;
 
-		addr = dma_map_single(dev->dev, iter->data, iter->len,
+		addr = dma_map_single(dev->dma_dev, iter->data, iter->len,
 				      DMA_TO_DEVICE);
-		if (unlikely(dma_mapping_error(dev->dev, addr)))
+		if (unlikely(dma_mapping_error(dev->dma_dev, addr)))
 			goto unmap;
 
 		tx_info.buf[n].addr = addr;
@@ -407,10 +407,10 @@ mt76_dma_tx_queue_skb(struct mt76_dev *dev, struct mt76_queue *q,
 		goto unmap;
 	}
 
-	dma_sync_single_for_cpu(dev->dev, t->dma_addr, dev->drv->txwi_size,
+	dma_sync_single_for_cpu(dev->dma_dev, t->dma_addr, dev->drv->txwi_size,
 				DMA_TO_DEVICE);
 	ret = dev->drv->tx_prepare_skb(dev, txwi, q->qid, wcid, sta, &tx_info);
-	dma_sync_single_for_device(dev->dev, t->dma_addr, dev->drv->txwi_size,
+	dma_sync_single_for_device(dev->dma_dev, t->dma_addr, dev->drv->txwi_size,
 				   DMA_TO_DEVICE);
 	if (ret < 0)
 		goto unmap;
@@ -420,7 +420,7 @@ mt76_dma_tx_queue_skb(struct mt76_dev *dev, struct mt76_queue *q,
 
 unmap:
 	for (n--; n > 0; n--)
-		dma_unmap_single(dev->dev, tx_info.buf[n].addr,
+		dma_unmap_single(dev->dma_dev, tx_info.buf[n].addr,
 				 tx_info.buf[n].len, DMA_TO_DEVICE);
 
 free:
@@ -465,8 +465,8 @@ mt76_dma_rx_fill(struct mt76_dev *dev, struct mt76_queue *q)
 		if (!buf)
 			break;
 
-		addr = dma_map_single(dev->dev, buf, len, DMA_FROM_DEVICE);
-		if (unlikely(dma_mapping_error(dev->dev, addr))) {
+		addr = dma_map_single(dev->dma_dev, buf, len, DMA_FROM_DEVICE);
+		if (unlikely(dma_mapping_error(dev->dma_dev, addr))) {
 			skb_free_frag(buf);
 			break;
 		}

drivers/net/wireless/mediatek/mt76/mac80211.c

@@ -545,6 +545,7 @@ mt76_alloc_device(struct device *pdev, unsigned int size,
 	dev->hw = hw;
 	dev->dev = pdev;
 	dev->drv = drv_ops;
+	dev->dma_dev = pdev;
 
 	phy = &dev->phy;
 	phy->dev = dev;

drivers/net/wireless/mediatek/mt76/mt76.h

@@ -698,6 +698,7 @@ struct mt76_dev {
 	const struct mt76_driver_ops *drv;
 	const struct mt76_mcu_ops *mcu_ops;
 	struct device *dev;
+	struct device *dma_dev;
 
 	struct mt76_mcu mcu;