hv_netvsc: Implement batching in send buffer

With this patch, we can send out multiple RNDIS data packets in one send buffer
slot and one VMBus message. It reduces the overhead associated with VMBus messages.

Signed-off-by: Haiyang Zhang <haiyangz@microsoft.com>
Reviewed-by: K. Y. Srinivasan <kys@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

4 files changed, 162 insertions, 47 deletions

diff --git a/drivers/net/hyperv/hyperv_net.h b/drivers/net/hyperv/hyperv_net.h
index 4815843a6019..384f057d6570 100644
--- a/drivers/net/hyperv/hyperv_net.h
+++ b/drivers/net/hyperv/hyperv_net.h
@@ -131,6 +131,7 @@ struct hv_netvsc_packet {
 
 	struct hv_device *device;
 	bool is_data_pkt;
+	bool xmit_more; /* from skb */
 	u16 vlan_tci;
 
 	u16 q_idx;
@@ -596,7 +597,16 @@ struct nvsp_message {
 
 #define VRSS_SEND_TAB_SIZE 16
 
-/* Per netvsc channel-specific */
+#define RNDIS_MAX_PKT_DEFAULT 8
+#define RNDIS_PKT_ALIGN_DEFAULT 8
+
+struct multi_send_data {
+	spinlock_t lock; /* protect struct multi_send_data */
+	struct hv_netvsc_packet *pkt; /* netvsc pkt pending */
+	u32 count; /* counter of batched packets */
+};
+
+/* Per netvsc device */
 struct netvsc_device {
 	struct hv_device *dev;
 
@@ -647,6 +657,10 @@ struct netvsc_device {
 	unsigned char *cb_buffer;
 	/* The sub channel callback buffer */
 	unsigned char *sub_cb_buf;
+
+	struct multi_send_data msd[NR_CPUS];
+	u32 max_pkt; /* max number of pkt in one send, e.g. 8 */
+	u32 pkt_align; /* alignment bytes, e.g. 8 */
 };
 
 /* NdisInitialize message */
diff --git a/drivers/net/hyperv/netvsc.c b/drivers/net/hyperv/netvsc.c
index 208eb05446ba..b81bd37d3afb 100644
--- a/drivers/net/hyperv/netvsc.c
+++ b/drivers/net/hyperv/netvsc.c
@@ -37,6 +37,7 @@ static struct netvsc_device *alloc_net_device(struct hv_device *device)
 {
 	struct netvsc_device *net_device;
 	struct net_device *ndev = hv_get_drvdata(device);
+	int i;
 
 	net_device = kzalloc(sizeof(struct netvsc_device), GFP_KERNEL);
 	if (!net_device)
@@ -53,6 +54,11 @@ static struct netvsc_device *alloc_net_device(struct hv_device *device)
 	net_device->destroy = false;
 	net_device->dev = device;
 	net_device->ndev = ndev;
+	net_device->max_pkt = RNDIS_MAX_PKT_DEFAULT;
+	net_device->pkt_align = RNDIS_PKT_ALIGN_DEFAULT;
+
+	for (i = 0; i < num_online_cpus(); i++)
+		spin_lock_init(&net_device->msd[i].lock);
 
 	hv_set_drvdata(device, net_device);
 	return net_device;
@@ -687,12 +693,23 @@ static u32 netvsc_get_next_send_section(struct netvsc_device *net_device)
 
 static u32 netvsc_copy_to_send_buf(struct netvsc_device *net_device,
 				   unsigned int section_index,
+				   u32 pend_size,
 				   struct hv_netvsc_packet *packet)
 {
 	char *start = net_device->send_buf;
-	char *dest = (start + (section_index * net_device->send_section_size));
+	char *dest = start + (section_index * net_device->send_section_size)
+		     + pend_size;
 	int i;
 	u32 msg_size = 0;
+	u32 padding = 0;
+	u32 remain = packet->total_data_buflen % net_device->pkt_align;
+
+	/* Add padding */
+	if (packet->is_data_pkt && packet->xmit_more && remain) {
+		padding = net_device->pkt_align - remain;
+		packet->rndis_msg->msg_len += padding;
+		packet->total_data_buflen += padding;
+	}
 
 	for (i = 0; i < packet->page_buf_cnt; i++) {
 		char *src = phys_to_virt(packet->page_buf[i].pfn << PAGE_SHIFT);
@@ -703,67 +720,48 @@ static u32 netvsc_copy_to_send_buf(struct netvsc_device *net_device,
 		msg_size += len;
 		dest += len;
 	}
+
+	if (padding) {
+		memset(dest, 0, padding);
+		msg_size += padding;
+	}
+
 	return msg_size;
 }
 
-int netvsc_send(struct hv_device *device,
-			struct hv_netvsc_packet *packet)
+static inline int netvsc_send_pkt(
+	struct hv_netvsc_packet *packet,
+	struct netvsc_device *net_device)
 {
-	struct netvsc_device *net_device;
-	int ret = 0;
-	struct nvsp_message sendMessage;
-	struct net_device *ndev;
-	struct vmbus_channel *out_channel = NULL;
-	u64 req_id;
-	unsigned int section_index = NETVSC_INVALID_INDEX;
-	u32 msg_size = 0;
-	struct sk_buff *skb = NULL;
+	struct nvsp_message nvmsg;
+	struct vmbus_channel *out_channel = packet->channel;
 	u16 q_idx = packet->q_idx;
+	struct net_device *ndev = net_device->ndev;
+	u64 req_id;
+	int ret;
 
-
-	net_device = get_outbound_net_device(device);
-	if (!net_device)
-		return -ENODEV;
-	ndev = net_device->ndev;
-
-	sendMessage.hdr.msg_type = NVSP_MSG1_TYPE_SEND_RNDIS_PKT;
+	nvmsg.hdr.msg_type = NVSP_MSG1_TYPE_SEND_RNDIS_PKT;
 	if (packet->is_data_pkt) {
 		/* 0 is RMC_DATA; */
-		sendMessage.msg.v1_msg.send_rndis_pkt.channel_type = 0;
+		nvmsg.msg.v1_msg.send_rndis_pkt.channel_type = 0;
 	} else {
 		/* 1 is RMC_CONTROL; */
-		sendMessage.msg.v1_msg.send_rndis_pkt.channel_type = 1;
+		nvmsg.msg.v1_msg.send_rndis_pkt.channel_type = 1;
 	}
 
-	/* Attempt to send via sendbuf */
-	if (packet->total_data_buflen < net_device->send_section_size) {
-		section_index = netvsc_get_next_send_section(net_device);
-		if (section_index != NETVSC_INVALID_INDEX) {
-			msg_size = netvsc_copy_to_send_buf(net_device,
-							   section_index,
-							   packet);
-			skb = (struct sk_buff *)
-			      (unsigned long)packet->send_completion_tid;
-			packet->page_buf_cnt = 0;
-		}
-	}
-	packet->send_buf_index = section_index;
-
-
-	sendMessage.msg.v1_msg.send_rndis_pkt.send_buf_section_index =
-		section_index;
-	sendMessage.msg.v1_msg.send_rndis_pkt.send_buf_section_size = msg_size;
+	nvmsg.msg.v1_msg.send_rndis_pkt.send_buf_section_index =
+		packet->send_buf_index;
+	if (packet->send_buf_index == NETVSC_INVALID_INDEX)
+		nvmsg.msg.v1_msg.send_rndis_pkt.send_buf_section_size = 0;
+	else
+		nvmsg.msg.v1_msg.send_rndis_pkt.send_buf_section_size =
+			packet->total_data_buflen;
 
 	if (packet->send_completion)
 		req_id = (ulong)packet;
 	else
 		req_id = 0;
 
-	out_channel = net_device->chn_table[packet->q_idx];
-	if (out_channel == NULL)
-		out_channel = device->channel;
-	packet->channel = out_channel;
-
 	if (out_channel->rescind)
 		return -ENODEV;
 
@@ -771,11 +769,12 @@ int netvsc_send(struct hv_device *device,
 		ret = vmbus_sendpacket_pagebuffer(out_channel,
 						  packet->page_buf,
 						  packet->page_buf_cnt,
-						  &sendMessage,
+						  &nvmsg,
 						  sizeof(struct nvsp_message),
 						  req_id);
 	} else {
-		ret = vmbus_sendpacket(out_channel, &sendMessage,
+		ret = vmbus_sendpacket(
+				out_channel, &nvmsg,
 				sizeof(struct nvsp_message),
 				req_id,
 				VM_PKT_DATA_INBAND,
@@ -809,6 +808,102 @@ int netvsc_send(struct hv_device *device,
 			   packet, ret);
 	}
 
+	return ret;
+}
+
+int netvsc_send(struct hv_device *device,
+		struct hv_netvsc_packet *packet)
+{
+	struct netvsc_device *net_device;
+	int ret = 0, m_ret = 0;
+	struct vmbus_channel *out_channel;
+	u16 q_idx = packet->q_idx;
+	u32 pktlen = packet->total_data_buflen, msd_len = 0;
+	unsigned int section_index = NETVSC_INVALID_INDEX;
+	struct sk_buff *skb = NULL;
+	unsigned long flag;
+	struct multi_send_data *msdp;
+	struct hv_netvsc_packet *msd_send = NULL, *cur_send = NULL;
+
+	net_device = get_outbound_net_device(device);
+	if (!net_device)
+		return -ENODEV;
+
+	out_channel = net_device->chn_table[q_idx];
+	if (!out_channel) {
+		out_channel = device->channel;
+		q_idx = 0;
+		packet->q_idx = 0;
+	}
+	packet->channel = out_channel;
+	packet->send_buf_index = NETVSC_INVALID_INDEX;
+
+	msdp = &net_device->msd[q_idx];
+
+	/* batch packets in send buffer if possible */
+	spin_lock_irqsave(&msdp->lock, flag);
+	if (msdp->pkt)
+		msd_len = msdp->pkt->total_data_buflen;
+
+	if (packet->is_data_pkt && msd_len > 0 &&
+	    msdp->count < net_device->max_pkt &&
+	    msd_len + pktlen + net_device->pkt_align <
+	    net_device->send_section_size) {
+		section_index = msdp->pkt->send_buf_index;
+
+	} else if (packet->is_data_pkt && pktlen + net_device->pkt_align <
+		   net_device->send_section_size) {
+		section_index = netvsc_get_next_send_section(net_device);
+		if (section_index != NETVSC_INVALID_INDEX) {
+				msd_send = msdp->pkt;
+				msdp->pkt = NULL;
+				msdp->count = 0;
+				msd_len = 0;
+		}
+	}
+
+	if (section_index != NETVSC_INVALID_INDEX) {
+		netvsc_copy_to_send_buf(net_device,
+					section_index, msd_len,
+					packet);
+		skb = (struct sk_buff *)
+		       (unsigned long)packet->send_completion_tid;
+
+		packet->page_buf_cnt = 0;
+		packet->send_buf_index = section_index;
+		packet->total_data_buflen += msd_len;
+
+		kfree(msdp->pkt);
+		if (packet->xmit_more) {
+			msdp->pkt = packet;
+			msdp->count++;
+		} else {
+			cur_send = packet;
+			msdp->pkt = NULL;
+			msdp->count = 0;
+		}
+	} else {
+		msd_send = msdp->pkt;
+		msdp->pkt = NULL;
+		msdp->count = 0;
+		cur_send = packet;
+	}
+
+	spin_unlock_irqrestore(&msdp->lock, flag);
+
+	if (msd_send) {
+		m_ret = netvsc_send_pkt(msd_send, net_device);
+
+		if (m_ret != 0) {
+			netvsc_free_send_slot(net_device,
+					      msd_send->send_buf_index);
+			kfree(msd_send);
+		}
+	}
+
+	if (cur_send)
+		ret = netvsc_send_pkt(cur_send, net_device);
+
 	if (ret != 0) {
 		if (section_index != NETVSC_INVALID_INDEX)
 			netvsc_free_send_slot(net_device, section_index);
diff --git a/drivers/net/hyperv/netvsc_drv.c b/drivers/net/hyperv/netvsc_drv.c
index a06bd6614007..0c998186039e 100644
--- a/drivers/net/hyperv/netvsc_drv.c
+++ b/drivers/net/hyperv/netvsc_drv.c
@@ -413,6 +413,8 @@ static int netvsc_start_xmit(struct sk_buff *skb, struct net_device *net)
 		return NETDEV_TX_OK;
 	}
 
+	packet->xmit_more = skb->xmit_more;
+
 	packet->vlan_tci = skb->vlan_tci;
 
 	packet->q_idx = skb_get_queue_mapping(skb);
diff --git a/drivers/net/hyperv/rndis_filter.c b/drivers/net/hyperv/rndis_filter.c
index ca81de04bc76..a2b185a58075 100644
--- a/drivers/net/hyperv/rndis_filter.c
+++ b/drivers/net/hyperv/rndis_filter.c
@@ -237,6 +237,7 @@ static int rndis_filter_send_request(struct rndis_device *dev,
 	}
 
 	packet->send_completion = NULL;
+	packet->xmit_more = false;
 
 	ret = netvsc_send(dev->net_dev->dev, packet);
 	return ret;
@@ -855,6 +856,7 @@ static int rndis_filter_init_device(struct rndis_device *dev)
 	u32 status;
 	int ret;
 	unsigned long t;
+	struct netvsc_device *nvdev = dev->net_dev;
 
 	request = get_rndis_request(dev, RNDIS_MSG_INIT,
 			RNDIS_MESSAGE_SIZE(struct rndis_initialize_request));
@@ -889,6 +891,8 @@ static int rndis_filter_init_device(struct rndis_device *dev)
 	status = init_complete->status;
 	if (status == RNDIS_STATUS_SUCCESS) {
 		dev->state = RNDIS_DEV_INITIALIZED;
+		nvdev->max_pkt = init_complete->max_pkt_per_msg;
+		nvdev->pkt_align = 1 << init_complete->pkt_alignment_factor;
 		ret = 0;
 	} else {
 		dev->state = RNDIS_DEV_UNINITIALIZED;