[PATCH net-next 6/6] net: mvneta: Improve the buffer allocation method for SWBM

[Gregory CLEMENT]

From: Yelena Krivosheev <yelena@xxxxxxxxxxx>

With system having a small memory (around 256MB), the state "cannot
allocate memory to refill with new buffer" is reach pretty quickly.

By this patch we changed buffer allocation method to a better handling of
this use case by avoiding memory allocation issues.

Signed-off-by: Yelena Krivosheev <yelena@xxxxxxxxxxx>
[gregory: extract from a larger patch]
Signed-off-by: Gregory CLEMENT <gregory.clement@xxxxxxxxxxx>
---
 drivers/net/ethernet/marvell/mvneta.c | 281 +++++++++++++++++---------
 1 file changed, 183 insertions(+), 98 deletions(-)

diff --git a/drivers/net/ethernet/marvell/mvneta.c b/drivers/net/ethernet/marvell/mvneta.c
index 8fc4be238083..806e78700e50 100644
--- a/drivers/net/ethernet/marvell/mvneta.c
+++ b/drivers/net/ethernet/marvell/mvneta.c
@@ -479,7 +479,10 @@ struct mvneta_port {
 #define MVNETA_RXD_ERR_RESOURCE		(BIT(17) | BIT(18))
 #define MVNETA_RXD_ERR_CODE_MASK	(BIT(17) | BIT(18))
 #define MVNETA_RXD_L3_IP4		BIT(25)
-#define MVNETA_RXD_FIRST_LAST_DESC	(BIT(26) | BIT(27))
+#define MVNETA_RXD_LAST_DESC		BIT(26)
+#define MVNETA_RXD_FIRST_DESC		BIT(27)
+#define MVNETA_RXD_FIRST_LAST_DESC	(MVNETA_RXD_FIRST_DESC | \
+					 MVNETA_RXD_LAST_DESC)
 #define MVNETA_RXD_L4_CSUM_OK		BIT(30)
 
 #if defined(__LITTLE_ENDIAN)
@@ -607,6 +610,14 @@ struct mvneta_rx_queue {
 	/* Index of the next RX DMA descriptor to process */
 	int next_desc_to_proc;
 
+	/* Index of first RX DMA descriptor to refill */
+	int first_to_refill;
+	u32 refill_num;
+
+	/* pointer to uncomplete skb buffer */
+	struct sk_buff *skb;
+	int left_size;
+
 	/* error counters */
 	u32 skb_alloc_err;
 	u32 refill_err;
@@ -622,6 +633,7 @@ static int txq_number = 8;
 static int rxq_def;
 
 static int rx_copybreak __read_mostly = 256;
+static int rx_header_size __read_mostly = 128;
 
 /* HW BM need that each port be identify by a unique ID */
 static int global_port_id;
@@ -1685,13 +1697,6 @@ static void mvneta_rx_error(struct mvneta_port *pp,
 {
 	u32 status = rx_desc->status;
 
-	if (!mvneta_rxq_desc_is_first_last(status)) {
-		netdev_err(pp->dev,
-			   "bad rx status %08x (buffer oversize), size=%d\n",
-			   status, rx_desc->data_size);
-		return;
-	}
-
 	switch (status & MVNETA_RXD_ERR_CODE_MASK) {
 	case MVNETA_RXD_ERR_CRC:
 		netdev_err(pp->dev, "bad rx status %08x (crc error), size=%d\n",
@@ -1879,6 +1884,8 @@ static void mvneta_rxq_drop_pkts(struct mvneta_port *pp,
 	for (i = 0; i < rxq->size; i++) {
 		struct mvneta_rx_desc *rx_desc = rxq->descs + i;
 		void *data = rxq->buf_virt_addr[i];
+		if (!data || !(rx_desc->buf_phys_addr))
+			continue;
 
 		dma_unmap_single(pp->dev->dev.parent, rx_desc->buf_phys_addr,
 				 MVNETA_RX_BUF_SIZE(pp->pkt_size), DMA_FROM_DEVICE);
@@ -1886,117 +1893,183 @@ static void mvneta_rxq_drop_pkts(struct mvneta_port *pp,
 	}
 }
 
+static inline
+int mvneta_rx_refill_queue(struct mvneta_port *pp, struct mvneta_rx_queue *rxq)
+{
+	struct mvneta_rx_desc *rx_desc;
+	int curr_desc = rxq->first_to_refill;
+	int i;
+
+	for (i = 0; (i < rxq->refill_num) && (i < 64); i++) {
+		rx_desc = rxq->descs + curr_desc;
+		if (!(rx_desc->buf_phys_addr)) {
+			if (mvneta_rx_refill(pp, rx_desc, rxq, GFP_ATOMIC)) {
+				pr_err("Can't refill queue %d. Done %d from %d\n",
+				       rxq->id, i, rxq->refill_num);
+				rxq->refill_err++;
+				break;
+			}
+		}
+		curr_desc = MVNETA_QUEUE_NEXT_DESC(rxq, curr_desc);
+	}
+	rxq->refill_num -= i;
+	rxq->first_to_refill = curr_desc;
+
+	return i;
+}
+
 /* Main rx processing when using software buffer management */
-static int mvneta_rx_swbm(struct mvneta_port *pp, int rx_todo,
+static int mvneta_rx_swbm(struct mvneta_port *pp, int budget,
 			  struct mvneta_rx_queue *rxq)
 {
 	struct mvneta_pcpu_port *port = this_cpu_ptr(pp->ports);
 	struct net_device *dev = pp->dev;
-	int rx_done;
+	int rx_todo, rx_proc;
+	int refill = 0;
 	u32 rcvd_pkts = 0;
 	u32 rcvd_bytes = 0;
 
 	/* Get number of received packets */
-	rx_done = mvneta_rxq_busy_desc_num_get(pp, rxq);
-
-	if (rx_todo > rx_done)
-		rx_todo = rx_done;
-
-	rx_done = 0;
+	rx_todo = mvneta_rxq_busy_desc_num_get(pp, rxq);
+	rx_proc = 0;
 
 	/* Fairness NAPI loop */
-	while (rx_done < rx_todo) {
+	while ((rcvd_pkts < budget) && (rx_proc < rx_todo)) {
 		struct mvneta_rx_desc *rx_desc = mvneta_rxq_next_desc_get(rxq);
-		struct sk_buff *skb;
 		unsigned char *data;
 		struct page *page;
 		dma_addr_t phys_addr;
-		u32 rx_status, frag_size;
-		int rx_bytes, err, index;
+		u32 rx_status, index;
+		int rx_bytes, skb_size, copy_size;
+		int frag_num, frag_size, frag_offset;
 
-		rx_done++;
-		rx_status = rx_desc->status;
-		rx_bytes = rx_desc->data_size - (ETH_FCS_LEN + MVNETA_MH_SIZE);
 		index = rx_desc - rxq->descs;
 		page = (struct page *)rxq->buf_virt_addr[index];
 		data = page_address(page);
 		/* Prefetch header */
 		prefetch(data);
-		phys_addr = rx_desc->buf_phys_addr - pp->rx_offset_correction;
 
-		if (!mvneta_rxq_desc_is_first_last(rx_status) ||
-		    (rx_status & MVNETA_RXD_ERR_SUMMARY)) {
-			mvneta_rx_error(pp, rx_desc);
-err_drop_frame:
-			dev->stats.rx_errors++;
-			/* leave the descriptor untouched */
-			continue;
-		}
+		phys_addr = rx_desc->buf_phys_addr;
+		rx_status = rx_desc->status;
+		rx_proc++;
+		rxq->refill_num++;
+
+		if (rx_status & MVNETA_RXD_FIRST_DESC) {
+			/* Check errors only for FIRST descriptor */
+			if (rx_status & MVNETA_RXD_ERR_SUMMARY) {
+				mvneta_rx_error(pp, rx_desc);
+				dev->stats.rx_errors++;
+				/* leave the descriptor untouched */
+				continue;
+			}
+			rx_bytes = rx_desc->data_size -
+				   (ETH_FCS_LEN + MVNETA_MH_SIZE);
 
-		if (rx_bytes <= rx_copybreak) {
-		/* better copy a small frame and not unmap the DMA region */
-			skb = netdev_alloc_skb_ip_align(dev, rx_bytes);
-			if (unlikely(!skb)) {
+			/* Allocate small skb for each new packet */
+			skb_size = max(rx_copybreak, rx_header_size);
+			rxq->skb = netdev_alloc_skb_ip_align(dev, skb_size);
+			if (unlikely(!rxq->skb)) {
 				netdev_err(dev,
 					   "Can't allocate skb on queue %d\n",
 					   rxq->id);
+				dev->stats.rx_dropped++;
 				rxq->skb_alloc_err++;
-				goto err_drop_frame;
+				continue;
 			}
+			copy_size = min(skb_size, rx_bytes);
+
+			/* Copy data from buffer to SKB, skip Marvell header */
+			memcpy(rxq->skb->data, data + MVNETA_MH_SIZE,
+			       copy_size);
+			skb_put(rxq->skb, copy_size);
+			rxq->left_size = rx_bytes - copy_size;
+
+			mvneta_rx_csum(pp, rx_status, rxq->skb);
+			if (rxq->left_size == 0) {
+				int size = copy_size + MVNETA_MH_SIZE;
+
+				dma_sync_single_range_for_cpu(dev->dev.parent,
+							      phys_addr, 0,
+							      size,
+							      DMA_FROM_DEVICE);
+
+				/* leave the descriptor and buffer untouched */
+			} else {
+				/* refill descriptor with new buffer later */
+				rx_desc->buf_phys_addr = 0;
+
+				frag_num = 0;
+				frag_offset = copy_size + MVNETA_MH_SIZE;
+				frag_size = min(rxq->left_size,
+						(int)(PAGE_SIZE - frag_offset));
+				skb_add_rx_frag(rxq->skb, frag_num, page,
+						frag_offset, frag_size,
+						PAGE_SIZE);
+				dma_unmap_single(dev->dev.parent, phys_addr,
+						 PAGE_SIZE, DMA_FROM_DEVICE);
+				rxq->left_size -= frag_size;
+			}
+		} else {
+			/* Middle or Last descriptor */
+			if (unlikely(!rxq->skb)) {
+				pr_debug("no skb for rx_status 0x%x\n",
+					 rx_status);
+				continue;
+			}
+			if (!rxq->left_size) {
+				/* last descriptor has only FCS */
+				/* and can be discarded */
+				dma_sync_single_range_for_cpu(dev->dev.parent,
+							      phys_addr, 0,
+							      ETH_FCS_LEN,
+							      DMA_FROM_DEVICE);
+				/* leave the descriptor and buffer untouched */
+			} else {
+				/* refill descriptor with new buffer later */
+				rx_desc->buf_phys_addr = 0;
+
+				frag_num = skb_shinfo(rxq->skb)->nr_frags;
+				frag_offset = 0;
+				frag_size = min(rxq->left_size,
+						(int)(PAGE_SIZE - frag_offset));
+				skb_add_rx_frag(rxq->skb, frag_num, page,
+						frag_offset, frag_size,
+						PAGE_SIZE);
+
+				dma_unmap_single(dev->dev.parent, phys_addr,
+						 PAGE_SIZE,
+						 DMA_FROM_DEVICE);
+
+				rxq->left_size -= frag_size;
+			}
+		} /* Middle or Last descriptor */
 
-			dma_sync_single_range_for_cpu(dev->dev.parent,
-						      phys_addr,
-						      MVNETA_MH_SIZE + NET_SKB_PAD,
-						      rx_bytes,
-						      DMA_FROM_DEVICE);
-			skb_put_data(skb, data + MVNETA_MH_SIZE + NET_SKB_PAD,
-				     rx_bytes);
-
-			skb->protocol = eth_type_trans(skb, dev);
-			mvneta_rx_csum(pp, rx_status, skb);
-			napi_gro_receive(&port->napi, skb);
-
-			rcvd_pkts++;
-			rcvd_bytes += rx_bytes;
-
-			/* leave the descriptor and buffer untouched */
+		if (!(rx_status & MVNETA_RXD_LAST_DESC))
+			/* no last descriptor this time */
 			continue;
-		}
 
-		/* Refill processing */
-		err = mvneta_rx_refill(pp, rx_desc, rxq, GFP_KERNEL);
-		if (err) {
-			netdev_err(dev, "Linux processing - Can't refill\n");
-			rxq->refill_err++;
-			goto err_drop_frame;
+		if (rxq->left_size) {
+			pr_err("get last desc, but left_size (%d) != 0\n",
+			       rxq->left_size);
+			dev_kfree_skb_any(rxq->skb);
+			rxq->left_size = 0;
+			rxq->skb = NULL;
+			continue;
 		}
-
-		frag_size = pp->frag_size;
-
-		skb = build_skb(data, frag_size > PAGE_SIZE ? 0 : frag_size);
-
-		/* After refill old buffer has to be unmapped regardless
-		 * the skb is successfully built or not.
-		 */
-		dma_unmap_single(dev->dev.parent, phys_addr,
-				 MVNETA_RX_BUF_SIZE(pp->pkt_size),
-				 DMA_FROM_DEVICE);
-
-		if (!skb)
-			goto err_drop_frame;
-
 		rcvd_pkts++;
-		rcvd_bytes += rx_bytes;
+		rcvd_bytes += rxq->skb->len;
 
 		/* Linux processing */
-		skb_reserve(skb, MVNETA_MH_SIZE + NET_SKB_PAD);
-		skb_put(skb, rx_bytes);
-
-		skb->protocol = eth_type_trans(skb, dev);
+		rxq->skb->protocol = eth_type_trans(rxq->skb, dev);
 
-		mvneta_rx_csum(pp, rx_status, skb);
+		if (dev->features & NETIF_F_GRO)
+			napi_gro_receive(&port->napi, rxq->skb);
+		else
+			netif_receive_skb(rxq->skb);
 
-		napi_gro_receive(&port->napi, skb);
+		/* clean uncomplete skb pointer in queue */
+		rxq->skb = NULL;
+		rxq->left_size = 0;
 	}
 
 	if (rcvd_pkts) {
@@ -2008,10 +2081,13 @@ static int mvneta_rx_swbm(struct mvneta_port *pp, int rx_todo,
 		u64_stats_update_end(&stats->syncp);
 	}
 
+	/* return some buffers to hardware queue, one at a time is too slow */
+	refill = mvneta_rx_refill_queue(pp, rxq);
+
 	/* Update rxq management counters */
-	mvneta_rxq_desc_num_update(pp, rxq, rx_done, rx_done);
+	mvneta_rxq_desc_num_update(pp, rxq, rx_proc, refill);
 
-	return rx_done;
+	return rcvd_pkts;
 }
 
 /* Main rx processing when using hardware buffer management */
@@ -2818,21 +2894,23 @@ static void mvneta_rxq_hw_init(struct mvneta_port *pp,
 	mvreg_write(pp, MVNETA_RXQ_BASE_ADDR_REG(rxq->id), rxq->descs_phys);
 	mvreg_write(pp, MVNETA_RXQ_SIZE_REG(rxq->id), rxq->size);
 
-	/* Set Offset */
-	mvneta_rxq_offset_set(pp, rxq, NET_SKB_PAD - pp->rx_offset_correction);
-
 	/* Set coalescing pkts and time */
 	mvneta_rx_pkts_coal_set(pp, rxq, rxq->pkts_coal);
 	mvneta_rx_time_coal_set(pp, rxq, rxq->time_coal);
 
 	if (!pp->bm_priv) {
-		/* Fill RXQ with buffers from RX pool */
-		mvneta_rxq_buf_size_set(pp, rxq,
-					MVNETA_RX_BUF_SIZE(pp->pkt_size));
+		/* Set Offset */
+		mvneta_rxq_offset_set(pp, rxq, 0);
+		mvneta_rxq_buf_size_set(pp, rxq, pp->frag_size);
 		mvneta_rxq_bm_disable(pp, rxq);
 		mvneta_rxq_fill(pp, rxq, rxq->size);
 	} else {
+		/* Set Offset */
+		mvneta_rxq_offset_set(pp, rxq,
+				      NET_SKB_PAD - pp->rx_offset_correction);
+
 		mvneta_rxq_bm_enable(pp, rxq);
+		/* Fill RXQ with buffers from RX pool */
 		mvneta_rxq_long_pool_set(pp, rxq);
 		mvneta_rxq_short_pool_set(pp, rxq);
 		mvneta_rxq_non_occup_desc_add(pp, rxq, rxq->size);
@@ -2861,6 +2939,9 @@ static void mvneta_rxq_deinit(struct mvneta_port *pp,
 {
 	mvneta_rxq_drop_pkts(pp, rxq);
 
+	if (rxq->skb)
+		dev_kfree_skb_any(rxq->skb);
+
 	if (rxq->descs)
 		dma_free_coherent(pp->dev->dev.parent,
 				  rxq->size * MVNETA_DESC_ALIGNED_SIZE,
@@ -2871,6 +2952,10 @@ static void mvneta_rxq_deinit(struct mvneta_port *pp,
 	rxq->last_desc         = 0;
 	rxq->next_desc_to_proc = 0;
 	rxq->descs_phys        = 0;
+	rxq->first_to_refill   = 0;
+	rxq->refill_num        = 0;
+	rxq->skb               = NULL;
+	rxq->left_size         = 0;
 }
 
 static int mvneta_txq_sw_init(struct mvneta_port *pp,
@@ -4356,14 +4441,6 @@ static int mvneta_probe(struct platform_device *pdev)
 	pp->dn = dn;
 
 	pp->rxq_def = rxq_def;
-
-	/* Set RX packet offset correction for platforms, whose
-	 * NET_SKB_PAD, exceeds 64B. It should be 64B for 64-bit
-	 * platforms and 0B for 32-bit ones.
-	 */
-	pp->rx_offset_correction =
-		max(0, NET_SKB_PAD - MVNETA_RX_PKT_OFFSET_CORRECTION);
-
 	pp->indir[0] = rxq_def;
 
 	/* Get special SoC configurations */
@@ -4451,6 +4528,7 @@ static int mvneta_probe(struct platform_device *pdev)
 	SET_NETDEV_DEV(dev, &pdev->dev);
 
 	pp->id = global_port_id++;
+	pp->rx_offset_correction = 0; /* not relevant for SW BM */
 
 	/* Obtain access to BM resources if enabled and already initialized */
 	bm_node = of_parse_phandle(dn, "buffer-manager", 0);
@@ -4465,6 +4543,13 @@ static int mvneta_probe(struct platform_device *pdev)
 				pp->bm_priv = NULL;
 			}
 		}
+		/* Set RX packet offset correction for platforms, whose
+		 * NET_SKB_PAD, exceeds 64B. It should be 64B for 64-bit
+		 * platforms and 0B for 32-bit ones.
+		 */
+		pp->rx_offset_correction = max(0,
+					       NET_SKB_PAD -
+					       MVNETA_RX_PKT_OFFSET_CORRECTION);
 	}
 	of_node_put(bm_node);
 
-- 
2.17.1