Merge branch 'hv_netvsc_linearize'

Vitaly Kuznetsov says:

====================
hv_netvsc: linearize SKBs bigger than MAX_PAGE_BUFFER_COUNT-2 pages

This patch series fixes the same issue which was fixed in Xen with commit
97a6d1bb2b658ac85ed88205ccd1ab809899884d ("xen-netfront: Fix handling packets on
compound pages with skb_linearize").

It is relatively easy to create a packet which is small in size but occupies
more than 30 (MAX_PAGE_BUFFER_COUNT-2) pages. Here is a kernel-mode reproducer
which tries sending a packet with only 34 bytes of payload (but on 34 pages)
and fails:

static int __init sendfb_init(void)
{
	struct socket *sock;
	int i, ret;
	struct sockaddr_in in4_addr = { 0 };
	struct page *pages[17];
	unsigned long flags;

	ret = sock_create_kern(AF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
	if (ret) {
		pr_err("failed to create socket: %d!\n", ret);
		return ret;
	}

	in4_addr.sin_family = AF_INET;
	/* www.google.com, 74.125.133.99 */
	in4_addr.sin_addr.s_addr = cpu_to_be32(0x4a7d8563);
	in4_addr.sin_port = cpu_to_be16(80);

	ret = sock->ops->connect(sock, (struct sockaddr *)&in4_addr, sizeof(in4_addr), 0);
	if (ret) {
		pr_err("failed to connect: %d!\n", ret);
		return ret;
	}

	/* We can send up to 17 frags */
	flags = MSG_MORE;
	for (i = 0; i < 17; i++) {
		if (i == 16)
			flags = MSG_EOR;
		pages[i] = alloc_pages(GFP_KERNEL | __GFP_COMP, 1);
		if (!pages[i]) {
			pr_err("out of memory!");
			goto free_pages;
		}
		sock->ops->sendpage(sock, pages[i], PAGE_SIZE -1, 2, flags);
	}

free_pages:
	for (; i > 0; i--)
		__free_pages(pages[i - 1], 1);

	printk("sendfb_init: test done\n");
        return -1;
}

module_init(sendfb_init);

MODULE_LICENSE("GPL");

A try to load such module results in multiple
'kernel: hv_netvsc vmbus_15 eth0: Packet too big: 100' messages as all retries
fail as well. It should also be possible to trigger the issue from userspace, I
expect e.g. NFS under heavy load to get stuck sometimes.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/hyperv/netvsc_drv.c

@@ -370,35 +370,49 @@ not_ip:
 static int netvsc_start_xmit(struct sk_buff *skb, struct net_device *net)
 {
 	struct net_device_context *net_device_ctx = netdev_priv(net);
-	struct hv_netvsc_packet *packet;
+	struct hv_netvsc_packet *packet = NULL;
 	int ret;
 	unsigned int num_data_pgs;
 	struct rndis_message *rndis_msg;
 	struct rndis_packet *rndis_pkt;
 	u32 rndis_msg_size;
 	bool isvlan;
+	bool linear = false;
 	struct rndis_per_packet_info *ppi;
 	struct ndis_tcp_ip_checksum_info *csum_info;
 	struct ndis_tcp_lso_info *lso_info;
 	int  hdr_offset;
 	u32 net_trans_info;
 	u32 hash;
-	u32 skb_length = skb->len;
-	u32 head_room = skb_headroom(skb);
+	u32 skb_length;
+	u32 head_room;
 	u32 pkt_sz;
 	struct hv_page_buffer page_buf[MAX_PAGE_BUFFER_COUNT];
 
 
 	/* We will atmost need two pages to describe the rndis
 	 * header. We can only transmit MAX_PAGE_BUFFER_COUNT number
-	 * of pages in a single packet.
+	 * of pages in a single packet. If skb is scattered around
+	 * more pages we try linearizing it.
 	 */
+
+check_size:
+	skb_length = skb->len;
+	head_room = skb_headroom(skb);
 	num_data_pgs = netvsc_get_slots(skb) + 2;
-	if (num_data_pgs > MAX_PAGE_BUFFER_COUNT) {
-		netdev_err(net, "Packet too big: %u\n", skb->len);
-		dev_kfree_skb(skb);
-		net->stats.tx_dropped++;
-		return NETDEV_TX_OK;
+	if (num_data_pgs > MAX_PAGE_BUFFER_COUNT && linear) {
+		net_alert_ratelimited("packet too big: %u pages (%u bytes)\n",
+				      num_data_pgs, skb->len);
+		ret = -EFAULT;
+		goto drop;
+	} else if (num_data_pgs > MAX_PAGE_BUFFER_COUNT) {
+		if (skb_linearize(skb)) {
+			net_alert_ratelimited("failed to linearize skb\n");
+			ret = -ENOMEM;
+			goto drop;
+		}
+		linear = true;
+		goto check_size;
 	}
 
 	pkt_sz = sizeof(struct hv_netvsc_packet) + RNDIS_AND_PPI_SIZE;
@@ -408,9 +422,8 @@ static int netvsc_start_xmit(struct sk_buff *skb, struct net_device *net)
 		if (!packet) {
 			/* out of memory, drop packet */
 			netdev_err(net, "unable to alloc hv_netvsc_packet\n");
-			dev_kfree_skb(skb);
-			net->stats.tx_dropped++;
-			return NETDEV_TX_OK;
+			ret = -ENOMEM;
+			goto drop;
 		}
 		packet->part_of_skb = false;
 	} else {
@@ -574,7 +587,7 @@ drop:
 		net->stats.tx_bytes += skb_length;
 		net->stats.tx_packets++;
 	} else {
-		if (!packet->part_of_skb)
+		if (packet && !packet->part_of_skb)
 			kfree(packet);
 		if (ret != -EAGAIN) {
 			dev_kfree_skb_any(skb);