openwrt-redmi-ax3000/package/kernel/qca-nss-dp/patches/0101-WiP-use-standard-DMA-api.patch

From 5b9d0813e457881f4970d330dad0e974abc30b83 Mon Sep 17 00:00:00 2001
From: hzy <hzyitc@outlook.com>
Date: Fri, 3 Mar 2023 15:44:17 +0000
Subject: [PATCH 1/3] WiP: use standard DMA api

Signed-off-by: hzy <hzyitc@outlook.com>
---
 hal/dp_ops/syn_gmac_dp/syn_dp_cfg_rx.c | 13 ++++--
 hal/dp_ops/syn_gmac_dp/syn_dp_cfg_tx.c |  2 +
 hal/dp_ops/syn_gmac_dp/syn_dp_rx.c     | 61 +++++++++++++++++++-------
 hal/dp_ops/syn_gmac_dp/syn_dp_tx.c     | 23 +++-------
 4 files changed, 63 insertions(+), 36 deletions(-)

diff --git a/hal/dp_ops/syn_gmac_dp/syn_dp_cfg_rx.c b/hal/dp_ops/syn_gmac_dp/syn_dp_cfg_rx.c
index 3153b2ef..2f23c4cb 100644
--- a/hal/dp_ops/syn_gmac_dp/syn_dp_cfg_rx.c
+++ b/hal/dp_ops/syn_gmac_dp/syn_dp_cfg_rx.c
@@ -34,13 +34,15 @@ static int syn_dp_cfg_rx_setup_desc_queue(struct syn_dp_info *dev_info)
 	/*
 	 * Allocate cacheable descriptors for Rx
 	 */
-	first_desc = kzalloc(sizeof(struct dma_desc_rx) * SYN_DP_RX_DESC_SIZE, GFP_KERNEL);
+	first_desc = dma_alloc_coherent(rx_info->dev,
+					sizeof(struct dma_desc_rx) * SYN_DP_RX_DESC_SIZE,
+					&dma_addr, GFP_KERNEL);
 	if (!first_desc) {
 		netdev_dbg(netdev, "Error in Rx Descriptor Memory allocation in Ring mode\n");
 		return -ENOMEM;
 	}
 
-	dev_info->rx_desc_dma_addr = (dma_addr_t)virt_to_phys(first_desc);
+	dev_info->rx_desc_dma_addr = dma_addr;
 	rx_info->rx_desc = first_desc;
 	syn_dp_gmac_rx_desc_init_ring(rx_info->rx_desc, SYN_DP_RX_DESC_SIZE);
 
@@ -99,6 +101,10 @@ void syn_dp_cfg_rx_cleanup_rings(struct syn_dp_info *dev_info)
 	for (i = 0; i < rx_info->busy_rx_desc_cnt; i++) {
 		rx_skb_index = (rx_skb_index + i) & SYN_DP_RX_DESC_MAX_INDEX;
 		rxdesc = rx_info->rx_desc;
+
+		dma_unmap_single(rx_info->dev, rxdesc->buffer1,
+				 rxdesc->length, DMA_FROM_DEVICE);
+
 		skb = rx_info->rx_buf_pool[rx_skb_index].skb;
 		if (unlikely(skb != NULL)) {
 			dev_kfree_skb_any(skb);
@@ -106,7 +112,8 @@ void syn_dp_cfg_rx_cleanup_rings(struct syn_dp_info *dev_info)
 		}
 	}
 
-	kfree(rx_info->rx_desc);
+	dma_free_coherent(rx_info->dev, (sizeof(struct dma_desc_rx) * SYN_DP_RX_DESC_SIZE),
+			  rx_info->rx_desc, dev_info->rx_desc_dma_addr);
 	rx_info->rx_desc = NULL;
 	dev_info->rx_desc_dma_addr = (dma_addr_t)0;
 }
diff --git a/hal/dp_ops/syn_gmac_dp/syn_dp_cfg_tx.c b/hal/dp_ops/syn_gmac_dp/syn_dp_cfg_tx.c
index bf5e19a0..284e8880 100644
--- a/hal/dp_ops/syn_gmac_dp/syn_dp_cfg_tx.c
+++ b/hal/dp_ops/syn_gmac_dp/syn_dp_cfg_tx.c
@@ -91,6 +91,8 @@ void syn_dp_cfg_tx_cleanup_rings(struct syn_dp_info *dev_info)
 		tx_skb_index = syn_dp_tx_inc_index(tx_skb_index, i);
 		txdesc = tx_info->tx_desc;
 
+		dma_unmap_single(tx_info->dev, txdesc->buffer1, txdesc->length, DMA_TO_DEVICE);
+
 		skb = tx_info->tx_buf_pool[tx_skb_index].skb;
 		if (unlikely(skb != NULL)) {
 			dev_kfree_skb_any(skb);
diff --git a/hal/dp_ops/syn_gmac_dp/syn_dp_rx.c b/hal/dp_ops/syn_gmac_dp/syn_dp_rx.c
index 1ddeb7d6..fff25538 100644
--- a/hal/dp_ops/syn_gmac_dp/syn_dp_rx.c
+++ b/hal/dp_ops/syn_gmac_dp/syn_dp_rx.c
@@ -73,17 +73,27 @@ static inline void syn_dp_rx_refill_one_desc(struct dma_desc_rx *rx_desc,
  */
 static inline void syn_dp_rx_inval_and_flush(struct syn_dp_info_rx *rx_info, uint32_t start, uint32_t end)
 {
+	int dma_size;
+
 	/*
-	 * Batched flush and invalidation of the rx descriptors
+	 * flush and invalidation of the rx descriptors
 	 */
 	if (end > start) {
-		dmac_flush_range_no_dsb((void *)&rx_info->rx_desc[start], (void *)&rx_info->rx_desc[end] + sizeof(struct dma_desc_rx));
+		dma_size = sizeof(struct dma_desc_rx) * (end - start + 1);
+		dma_sync_single_for_device(rx_info->dev,
+					   rx_info->rx_desc[start].buffer1,
+					   dma_size, DMA_FROM_DEVICE);
 	} else {
-		dmac_flush_range_no_dsb((void *)&rx_info->rx_desc[start], (void *)&rx_info->rx_desc[SYN_DP_RX_DESC_MAX_INDEX] + sizeof(struct dma_desc_rx));
-		dmac_flush_range_no_dsb((void *)&rx_info->rx_desc[0], (void *)&rx_info->rx_desc[end] + sizeof(struct dma_desc_rx));
+		dma_size = sizeof(struct dma_desc_rx) * (SYN_DP_RX_DESC_MAX_INDEX - start + 1);
+		dma_sync_single_for_device(rx_info->dev,
+					   rx_info->rx_desc[start].buffer1,
+					   dma_size, DMA_FROM_DEVICE);
+
+		dma_size = sizeof(struct dma_desc_rx) * (end + 1);
+		dma_sync_single_for_device(rx_info->dev,
+					   rx_info->rx_desc[0].buffer1,
+					   dma_size, DMA_FROM_DEVICE);
 	}
-
-	dsb(st);
 }
 
 /*
@@ -127,12 +137,15 @@ int syn_dp_rx_refill_page_mode(struct syn_dp_info_rx *rx_info)
 		/*
 		 * Get virtual address of allocated page.
 		 */
-		page_addr = page_address(pg);
-		dma_addr = (dma_addr_t)virt_to_phys(page_addr);
+		dma_addr = dma_map_page(rx_info->dev, pg, 0, rx_info->alloc_buf_len, DMA_FROM_DEVICE);
+		if (unlikely(dma_mapping_error(rx_info->dev, dma_addr))) {
+			dev_kfree_skb(skb);
+			netdev_dbg(netdev, "DMA mapping failed for empty buffer\n");
+			break;
+		}
 
 		skb_fill_page_desc(skb, 0, pg, 0, PAGE_SIZE);
 
-		dmac_inv_range_no_dsb(page_addr, (page_addr + PAGE_SIZE));
 		rx_refill_idx = rx_info->rx_refill_idx;
 		rx_desc = rx_info->rx_desc + rx_refill_idx;
 
@@ -181,8 +194,13 @@ int syn_dp_rx_refill(struct syn_dp_info_rx *rx_info)
 
 		skb_reserve(skb, SYN_DP_SKB_HEADROOM + NET_IP_ALIGN);
 
-		dma_addr = (dma_addr_t)virt_to_phys(skb->data);
-		dmac_inv_range_no_dsb((void *)skb->data, (void *)(skb->data + inval_len));
+		dma_addr = dma_map_single(rx_info->dev, skb->data, rx_info->alloc_buf_len, DMA_FROM_DEVICE);
+		if (unlikely(dma_mapping_error(rx_info->dev, dma_addr))) {
+			dev_kfree_skb(skb);
+			netdev_dbg(netdev, "DMA mapping failed for empty buffer\n");
+			break;
+		}
+
 		rx_refill_idx = rx_info->rx_refill_idx;
 		rx_desc = rx_info->rx_desc + rx_refill_idx;
 
@@ -381,6 +399,7 @@ int syn_dp_rx(struct syn_dp_info_rx *rx_info, int budget)
 	struct dma_desc_rx *rx_desc_next = NULL;
 	uint8_t *next_skb_ptr;
 	skb_frag_t *frag = NULL;
+	int dma_size;
 
 	busy = atomic_read((atomic_t *)&rx_info->busy_rx_desc_cnt);
 	if (unlikely(!busy)) {
@@ -408,10 +427,20 @@ int syn_dp_rx(struct syn_dp_info_rx *rx_info, int budget)
 	 */
 	end = syn_dp_rx_inc_index(rx_info->rx_idx, busy);
 	if (end > start) {
-		dmac_inv_range_no_dsb((void *)&rx_info->rx_desc[start], (void *)&rx_info->rx_desc[end] + sizeof(struct dma_desc_rx));
+		dma_size = sizeof(struct dma_desc_rx) * (end - start + 1);
+		dma_sync_single_for_cpu(rx_info->dev,
+					rx_info->rx_desc[start].buffer1,
+					dma_size, DMA_FROM_DEVICE);
 	} else {
-		dmac_inv_range_no_dsb((void *)&rx_info->rx_desc[start], (void *)&rx_info->rx_desc[SYN_DP_RX_DESC_MAX_INDEX] + sizeof(struct dma_desc_rx));
-		dmac_inv_range_no_dsb((void *)&rx_info->rx_desc[0], (void *)&rx_info->rx_desc[end] + sizeof(struct dma_desc_rx));
+		dma_size = sizeof(struct dma_desc_rx) * (SYN_DP_RX_DESC_MAX_INDEX - start + 1);
+		dma_sync_single_for_cpu(rx_info->dev,
+					rx_info->rx_desc[start].buffer1,
+					dma_size, DMA_FROM_DEVICE);
+
+		dma_size = sizeof(struct dma_desc_rx) * (end + 1);
+		dma_sync_single_for_cpu(rx_info->dev,
+					rx_info->rx_desc[0].buffer1,
+					dma_size, DMA_FROM_DEVICE);
 	}
 
 	dsb(st);
@@ -439,8 +468,8 @@ int syn_dp_rx(struct syn_dp_info_rx *rx_info, int budget)
 		 * speculative prefetch by CPU may have occurred.
 		 */
 		frame_length = syn_dp_gmac_get_rx_desc_frame_length(status);
-		dmac_inv_range((void *)rx_buf->map_addr_virt,
-			(void *)(((uint8_t *)rx_buf->map_addr_virt) + frame_length));
+		dma_sync_single_for_cpu(rx_info->dev, rx_desc->buffer1,
+					frame_length, DMA_FROM_DEVICE);
 		prefetch((void *)rx_buf->map_addr_virt);
 
 		rx_next_idx = syn_dp_rx_inc_index(rx_idx, 1);
diff --git a/hal/dp_ops/syn_gmac_dp/syn_dp_tx.c b/hal/dp_ops/syn_gmac_dp/syn_dp_tx.c
index c97e252b..6d4adb3f 100644
--- a/hal/dp_ops/syn_gmac_dp/syn_dp_tx.c
+++ b/hal/dp_ops/syn_gmac_dp/syn_dp_tx.c
@@ -104,9 +104,7 @@ static inline struct dma_desc_tx *syn_dp_tx_process_nr_frags(struct syn_dp_info_
 		BUG_ON(!length);
 #endif
 
-		dma_addr = (dma_addr_t)virt_to_phys(frag_addr);
-
-		dmac_clean_range_no_dsb(frag_addr, frag_addr + length);
+		dma_addr = dma_map_single(tx_info->dev, frag_addr, length, DMA_TO_DEVICE);
 
 		*total_length += length;
 		tx_desc = syn_dp_tx_set_desc_sg(tx_info, dma_addr, length, DESC_OWN_BY_DMA);
@@ -150,8 +148,7 @@ int syn_dp_tx_nr_frags(struct syn_dp_info_tx *tx_info, struct sk_buff *skb)
 	/*
 	 * Flush the dma for non-paged skb data
 	 */
-	dma_addr = (dma_addr_t)virt_to_phys(skb->data);
-	dmac_clean_range_no_dsb((void *)skb->data, (void *)(skb->data + length));
+	dma_addr = dma_map_single(tx_info->dev, skb->data, length, DMA_TO_DEVICE);
 
 	total_len = length;
 
@@ -256,12 +253,7 @@ int syn_dp_tx_frag_list(struct syn_dp_info_tx *tx_info, struct sk_buff *skb)
 		return NETDEV_TX_BUSY;
 	}
 
-	dma_addr = (dma_addr_t)virt_to_phys(skb->data);
-
-	/*
-	 * Flush the data area of the head skb
-	 */
-	dmac_clean_range_no_dsb((void *)skb->data, (void *)(skb->data + length));
+	dma_addr = dma_map_single(tx_info->dev, skb->data, length, DMA_TO_DEVICE);
 
 	total_len = length;
 
@@ -290,9 +282,7 @@ int syn_dp_tx_frag_list(struct syn_dp_info_tx *tx_info, struct sk_buff *skb)
 		BUG_ON(!length);
 #endif
 
-		dma_addr = (dma_addr_t)virt_to_phys(iter_skb->data);
-
-		dmac_clean_range_no_dsb((void *)iter_skb->data, (void *)(iter_skb->data + length));
+		dma_addr = dma_map_single(tx_info->dev, iter_skb->data, length, DMA_TO_DEVICE);
 
 		total_len += length;
 
@@ -445,6 +435,7 @@ int syn_dp_tx_complete(struct syn_dp_info_tx *tx_info, int budget)
 			break;
 		}
 
+		dma_unmap_single(tx_info->dev, desc->buffer1, desc->length, DMA_TO_DEVICE);
 
 		if (likely(status & DESC_TX_LAST)) {
 			tx_skb_index = syn_dp_tx_comp_index_get(tx_info);
@@ -571,9 +562,7 @@ int syn_dp_tx(struct syn_dp_info_tx *tx_info, struct sk_buff *skb)
 		return NETDEV_TX_BUSY;
 	}
 
-	dma_addr = (dma_addr_t)virt_to_phys(skb->data);
-
-	dmac_clean_range_no_dsb((void *)skb->data, (void *)(skb->data + skb->len));
+	dma_addr = dma_map_single(tx_info->dev, skb->data, skb->len, DMA_TO_DEVICE);
 
 	/*
 	 * Queue packet to the GMAC rings
-- 
2.40.1