qcacmn: Add workaround for TXDMA HW limitation

Per HW team's analysis, we find a TXDMA HW limitation:
ADDR0&0x1FFFFFFF8 should not equal ADDR1&0x1FFFFFFF8.
Otherwise, TXDMA will run into exception, which cause TX fail.
ADDR0: the address of last words in previous buffer;
ADDR1: the address of first words in next buffer;

We hit this limitation in NAT forward TSO jumbo case whose buffer
address of two fragments like below:
tso_frags = (
(length = 0x42, vaddr = 0xFFFFFFD0F5FA2F82, paddr = 0x1F5FA2F82),
(length = 0x05A8, vaddr = 0xFFFFFFD0F5FA2FC4, paddr = 0x1F5FA2FC4)
In this case, ADDR0 = 0x1F5FA2F82 + 0x42 -2 = 0x1F5FA2FC2,
ADDR1 = 0x1F5FA2FC4, then
ADDR0&0x1FFFFFFF8 = ADDR1&0x1FFFFFFF8.
To avoid this, shift server bytes for ADDR0.

Change-Id: Ib09afd7e4c14a33bb5347c68b602b360a9e36619
CRs-Fixed: 2585868

qdf/linux/src/qdf_nbuf.c

@@ -2800,6 +2800,59 @@ struct qdf_tso_cmn_seg_info_t {
 	uint32_t tcp_seq_num;
 };
 
+/**
+ * qdf_nbuf_adj_tso_frag() - adjustment for buffer address of tso fragment
+ *
+ * @skb: network buffer
+ *
+ * Return: byte offset length of 8 bytes aligned.
+ */
+#ifdef WAR_TXDMA_LIMITATION
+static uint8_t qdf_nbuf_adj_tso_frag(struct sk_buff *skb)
+{
+	uint32_t eit_hdr_len;
+	uint8_t *eit_hdr;
+	uint8_t byte_8_align_offset;
+
+	/*
+	 * Workaround for TXDMA HW limitation.
+	 * ADDR0&0x1FFFFFFF8 should not equal ADDR1&0x1FFFFFFF8.
+	 * Otherwise, TXDMA will run into exception, which cause TX fail.
+	 * ADDR0: the address of last words in previous buffer;
+	 * ADDR1: the address of first words in next buffer;
+	 * To avoid this, shift several bytes for ADDR0.
+	 */
+	eit_hdr = skb->data;
+	eit_hdr_len = (skb_transport_header(skb)
+		 - skb_mac_header(skb)) + tcp_hdrlen(skb);
+	byte_8_align_offset = ((unsigned long)(eit_hdr) + eit_hdr_len) & 0x7L;
+	if (qdf_unlikely(byte_8_align_offset)) {
+		TSO_DEBUG("%pK,Len %d %d",
+			  eit_hdr, eit_hdr_len, byte_8_align_offset);
+		if (unlikely(skb_headroom(skb) < byte_8_align_offset)) {
+			TSO_DEBUG("[%d]Insufficient headroom,[%pK],[%pK],[%d]",
+				  __LINE__, skb->head, skb->data,
+				 byte_8_align_offset);
+			return 0;
+		}
+		qdf_nbuf_push_head(skb, byte_8_align_offset);
+		qdf_mem_move(skb->data,
+			     skb->data + byte_8_align_offset,
+			     eit_hdr_len);
+		skb->len -= byte_8_align_offset;
+		skb->mac_header -= byte_8_align_offset;
+		skb->network_header -= byte_8_align_offset;
+		skb->transport_header -= byte_8_align_offset;
+	}
+	return byte_8_align_offset;
+}
+#else
+static uint8_t qdf_nbuf_adj_tso_frag(struct sk_buff *skb)
+{
+	return 0;
+}
+#endif
+
 /**
  * __qdf_nbuf_get_tso_cmn_seg_info() - get TSO common
  * information
@@ -2970,12 +3023,15 @@ uint32_t __qdf_nbuf_get_tso_info(qdf_device_t osdev, struct sk_buff *skb,
 	uint32_t skb_proc = skb->len; /* bytes of skb pending processing */
 	uint32_t tso_seg_size = skb_shinfo(skb)->gso_size;
 	int j = 0; /* skb fragment index */
+	uint8_t byte_8_align_offset;
 
 	memset(&tso_cmn_info, 0x0, sizeof(tso_cmn_info));
 	total_num_seg = tso_info->tso_num_seg_list;
 	curr_seg = tso_info->tso_seg_list;
 	total_num_seg->num_seg.tso_cmn_num_seg = 0;
 
+	byte_8_align_offset = qdf_nbuf_adj_tso_frag(skb);
+
 	if (qdf_unlikely(__qdf_nbuf_get_tso_cmn_seg_info(osdev,
 						skb, &tso_cmn_info))) {
 		qdf_warn("TSO: error getting common segment info");
@@ -2991,7 +3047,9 @@ uint32_t __qdf_nbuf_get_tso_info(qdf_device_t osdev, struct sk_buff *skb,
 	skb_proc -= tso_cmn_info.eit_hdr_len;
 
 	/* get the address to the next tso fragment */
-	tso_frag_vaddr = skb->data + tso_cmn_info.eit_hdr_len;
+	tso_frag_vaddr = skb->data +
+			 tso_cmn_info.eit_hdr_len +
+			 byte_8_align_offset;
 	/* get the length of the next tso fragment */
 	tso_frag_len = min(skb_frag_len, tso_seg_size);