openwrt/target/linux/generic/pending-5.15/704-05-v6.4-net-mvneta-allocate-TSO-header-DMA-memory-in-chunks.patch

From 58d50fb089da553023df5a05f5ae86feaacc7f24 Mon Sep 17 00:00:00 2001
From: "Russell King (Oracle)" <[email protected]>
Date: Mon, 3 Apr 2023 19:30:40 +0100
Subject: [PATCH 5/5] net: mvneta: allocate TSO header DMA memory in chunks

Now that we no longer need to check whether the DMA address is within
the TSO header DMA memory range for the queue, we can allocate the TSO
header DMA memory in chunks rather than one contiguous order-6 chunk,
which can stress the kernel's memory subsystems to allocate.

Instead, use order-1 (8k) allocations, which will result in 32 order-1
pages containing 32 TSO headers.

Signed-off-by: Russell King (Oracle) <[email protected]>
---
 drivers/net/ethernet/marvell/mvneta.c | 88 +++++++++++++++++++++------
 1 file changed, 70 insertions(+), 18 deletions(-)

--- a/drivers/net/ethernet/marvell/mvneta.c
+++ b/drivers/net/ethernet/marvell/mvneta.c
@@ -314,6 +314,15 @@
 
 #define MVNETA_MAX_SKB_DESCS (MVNETA_MAX_TSO_SEGS * 2 + MAX_SKB_FRAGS)
 
+/* The size of a TSO header page */
+#define MVNETA_TSO_PAGE_SIZE (2 * PAGE_SIZE)
+
+/* Number of TSO headers per page. This should be a power of 2 */
+#define MVNETA_TSO_PER_PAGE (MVNETA_TSO_PAGE_SIZE / TSO_HEADER_SIZE)
+
+/* Maximum number of TSO header pages */
+#define MVNETA_MAX_TSO_PAGES (MVNETA_MAX_TXD / MVNETA_TSO_PER_PAGE)
+
 /* descriptor aligned size */
 #define MVNETA_DESC_ALIGNED_SIZE	32
 
@@ -656,10 +665,10 @@ struct mvneta_tx_queue {
 	int next_desc_to_proc;
 
 	/* DMA buffers for TSO headers */
-	char *tso_hdrs;
+	char *tso_hdrs[MVNETA_MAX_TSO_PAGES];
 
 	/* DMA address of TSO headers */
-	dma_addr_t tso_hdrs_phys;
+	dma_addr_t tso_hdrs_phys[MVNETA_MAX_TSO_PAGES];
 
 	/* Affinity mask for CPUs*/
 	cpumask_t affinity_mask;
@@ -2592,24 +2601,71 @@ err_drop_frame:
 	return rx_done;
 }
 
+static void mvneta_free_tso_hdrs(struct mvneta_port *pp,
+				 struct mvneta_tx_queue *txq)
+{
+	struct device *dev = pp->dev->dev.parent;
+	int i;
+
+	for (i = 0; i < MVNETA_MAX_TSO_PAGES; i++) {
+		if (txq->tso_hdrs[i]) {
+			dma_free_coherent(dev, MVNETA_TSO_PAGE_SIZE,
+					  txq->tso_hdrs[i],
+					  txq->tso_hdrs_phys[i]);
+			txq->tso_hdrs[i] = NULL;
+		}
+	}
+}
+
+static int mvneta_alloc_tso_hdrs(struct mvneta_port *pp,
+				 struct mvneta_tx_queue *txq)
+{
+	struct device *dev = pp->dev->dev.parent;
+	int i, num;
+
+	num = DIV_ROUND_UP(txq->size, MVNETA_TSO_PER_PAGE);
+	for (i = 0; i < num; i++) {
+		txq->tso_hdrs[i] = dma_alloc_coherent(dev, MVNETA_TSO_PAGE_SIZE,
+						      &txq->tso_hdrs_phys[i],
+						      GFP_KERNEL);
+		if (!txq->tso_hdrs[i]) {
+			mvneta_free_tso_hdrs(pp, txq);
+			return -ENOMEM;
+		}
+	}
+
+	return 0;
+}
+
+static char *mvneta_get_tso_hdr(struct mvneta_tx_queue *txq, dma_addr_t *dma)
+{
+	int index, offset;
+
+	index = txq->txq_put_index / MVNETA_TSO_PER_PAGE;
+	offset = (txq->txq_put_index % MVNETA_TSO_PER_PAGE) * TSO_HEADER_SIZE;
+
+	*dma = txq->tso_hdrs_phys[index] + offset;
+
+	return txq->tso_hdrs[index] + offset;
+}
+
 static void mvneta_tso_put_hdr(struct sk_buff *skb, struct mvneta_tx_queue *txq,
 			       struct tso_t *tso, int size, bool is_last)
 {
-	int tso_offset, hdr_len = skb_transport_offset(skb) + tcp_hdrlen(skb);
+	int hdr_len = skb_transport_offset(skb) + tcp_hdrlen(skb);
 	struct mvneta_tx_buf *buf = &txq->buf[txq->txq_put_index];
 	struct mvneta_tx_desc *tx_desc;
+	dma_addr_t hdr_phys;
 	char *hdr;
 
-	tso_offset = txq->txq_put_index * TSO_HEADER_SIZE;
-
-	hdr = txq->tso_hdrs + tso_offset;
+	hdr = mvneta_get_tso_hdr(txq, &hdr_phys);
 	tso_build_hdr(skb, hdr, tso, size, is_last);
 
 	tx_desc = mvneta_txq_next_desc_get(txq);
 	tx_desc->data_size = hdr_len;
 	tx_desc->command = mvneta_skb_tx_csum(skb);
 	tx_desc->command |= MVNETA_TXD_F_DESC;
-	tx_desc->buf_phys_addr = txq->tso_hdrs_phys + tso_offset;
+	tx_desc->buf_phys_addr = hdr_phys;
 	buf->type = MVNETA_TYPE_TSO;
 	buf->skb = NULL;
 
@@ -3401,7 +3457,7 @@ static void mvneta_rxq_deinit(struct mvn
 static int mvneta_txq_sw_init(struct mvneta_port *pp,
 			      struct mvneta_tx_queue *txq)
 {
-	int cpu;
+	int cpu, err;
 
 	txq->size = pp->tx_ring_size;
 
@@ -3426,11 +3482,9 @@ static int mvneta_txq_sw_init(struct mvn
 		return -ENOMEM;
 
 	/* Allocate DMA buffers for TSO MAC/IP/TCP headers */
-	txq->tso_hdrs = dma_alloc_coherent(pp->dev->dev.parent,
-					   txq->size * TSO_HEADER_SIZE,
-					   &txq->tso_hdrs_phys, GFP_KERNEL);
-	if (!txq->tso_hdrs)
-		return -ENOMEM;
+	err = mvneta_alloc_tso_hdrs(pp, txq);
+	if (err)
+		return err;
 
 	/* Setup XPS mapping */
 	if (pp->neta_armada3700)
@@ -3482,10 +3536,7 @@ static void mvneta_txq_sw_deinit(struct
 
 	kfree(txq->buf);
 
-	if (txq->tso_hdrs)
-		dma_free_coherent(pp->dev->dev.parent,
-				  txq->size * TSO_HEADER_SIZE,
-				  txq->tso_hdrs, txq->tso_hdrs_phys);
+	mvneta_free_tso_hdrs(pp, txq);
 	if (txq->descs)
 		dma_free_coherent(pp->dev->dev.parent,
 				  txq->size * MVNETA_DESC_ALIGNED_SIZE,
@@ -3494,7 +3545,6 @@ static void mvneta_txq_sw_deinit(struct
 	netdev_tx_reset_queue(nq);
 
 	txq->buf               = NULL;
-	txq->tso_hdrs          = NULL;
 	txq->descs             = NULL;
 	txq->last_desc         = 0;
 	txq->next_desc_to_proc = 0;
@@ -5543,6 +5593,8 @@ static int __init mvneta_driver_init(voi
 {
 	int ret;
 
+	BUILD_BUG_ON_NOT_POWER_OF_2(MVNETA_TSO_PER_PAGE);
+
 	ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "net/mvneta:online",
 				      mvneta_cpu_online,
 				      mvneta_cpu_down_prepare);