[PATCH net-next v3] ibmvnic: Enable TSO support

[PATCH net-next v3] ibmvnic: Enable TSO support

[Thomas Falcon]

Enable TSO in the ibmvnic driver. Scatter-gather is also enabled,
though there currently is no support for scatter-gather in
vNIC-compatible hardware, resulting in forced linearization
of all fragmented SKB's. Though not ideal, given the throughput
improvement gained by enabling TSO, it has been decided
that this is an acceptable tradeoff.

The feature is also enabled by a module parameter.
This parameter is necessary because TSO can not easily be
enabled or disabled in firmware without reinitializing the driver.

CC: Nathan Fontenot <nfont@linux.vnet.ibm.com>
CC: John Allen <jallen@linux.vnet.ibm.com>
Signed-off-by: Thomas Falcon <tlfalcon@linux.vnet.ibm.com>
---
v3: Remove module parameter, TSO will be set through ethtool
    include ibmvnic_fix_features to manage scatter-gather
v2: Fix kbuild robot warning from module parameter init
---
 drivers/net/ethernet/ibm/ibmvnic.c | 42 +++++++++++++++++++++++++++++++++++---
 1 file changed, 39 insertions(+), 3 deletions(-)

diff --git a/drivers/net/ethernet/ibm/ibmvnic.c b/drivers/net/ethernet/ibm/ibmvnic.c
index abe2b6e..b1f4a95 100644
--- a/drivers/net/ethernet/ibm/ibmvnic.c
+++ b/drivers/net/ethernet/ibm/ibmvnic.c
@@ -149,6 +149,20 @@ struct ibmvnic_stat {
 	{"internal_mac_rx_errors", IBMVNIC_STAT_OFF(internal_mac_rx_errors)},
 };
 
+static netdev_features_t ibmvnic_fix_features(struct net_device *dev,
+					      netdev_features_t features)
+{
+	/* Scatter-gather is not currently supported by firmware.
+	 * It will only be enabled to support TSO operations.
+	 */
+	if (!(features & (NETIF_F_TSO | NETIF_F_TSO6)))
+		features &= ~NETIF_F_SG;
+	else
+		features |= NETIF_F_SG;
+
+	return features;
+}
+
 static long h_reg_sub_crq(unsigned long unit_address, unsigned long token,
 			  unsigned long length, unsigned long *number,
 			  unsigned long *irq)
@@ -1025,6 +1039,17 @@ static int ibmvnic_xmit(struct sk_buff *skb, struct net_device *netdev)
 		goto out;
 	}
 
+	/* All scatter-gather SKB's will be linearized for the time being, but
+	 * scatter-gather is only enabled if the user wishes to use TSO.
+	 */
+	if (skb_shinfo(skb)->nr_frags && __skb_linearize(skb)) {
+		dev_kfree_skb_any(skb);
+		tx_send_failed++;
+		tx_dropped++;
+		ret = NETDEV_TX_OK;
+		goto out;
+	}
+
 	tx_pool = &adapter->tx_pool[queue_num];
 	tx_scrq = adapter->tx_scrq[queue_num];
 	txq = netdev_get_tx_queue(netdev, skb_get_queue_mapping(skb));
@@ -1082,6 +1107,11 @@ static int ibmvnic_xmit(struct sk_buff *skb, struct net_device *netdev)
 		tx_crq.v1.flags1 |= IBMVNIC_TX_CHKSUM_OFFLOAD;
 		hdrs += 2;
 	}
+	if (skb_is_gso(skb)) {
+		tx_crq.v1.flags1 |= IBMVNIC_TX_LSO;
+		tx_crq.v1.mss = cpu_to_be16(skb_shinfo(skb)->gso_size);
+		hdrs += 2;
+	}
 	/* determine if l2/3/4 headers are sent to firmware */
 	if ((*hdrs >> 7) & 1 &&
 	    (skb->protocol == htons(ETH_P_IP) ||
@@ -1520,6 +1550,7 @@ static void ibmvnic_netpoll_controller(struct net_device *dev)
 	.ndo_set_mac_address	= ibmvnic_set_mac,
 	.ndo_validate_addr	= eth_validate_addr,
 	.ndo_tx_timeout		= ibmvnic_tx_timeout,
+	.ndo_fix_features	= ibmvnic_fix_features,
 #ifdef CONFIG_NET_POLL_CONTROLLER
 	.ndo_poll_controller	= ibmvnic_netpoll_controller,
 #endif
@@ -2629,10 +2660,10 @@ static void handle_query_ip_offload_rsp(struct ibmvnic_adapter *adapter)
 	adapter->ip_offload_ctrl.udp_ipv4_chksum = buf->udp_ipv4_chksum;
 	adapter->ip_offload_ctrl.tcp_ipv6_chksum = buf->tcp_ipv6_chksum;
 	adapter->ip_offload_ctrl.udp_ipv6_chksum = buf->udp_ipv6_chksum;
+	adapter->ip_offload_ctrl.large_tx_ipv4 = buf->large_tx_ipv4;
+	adapter->ip_offload_ctrl.large_tx_ipv6 = buf->large_tx_ipv6;
 
-	/* large_tx/rx disabled for now, additional features needed */
-	adapter->ip_offload_ctrl.large_tx_ipv4 = 0;
-	adapter->ip_offload_ctrl.large_tx_ipv6 = 0;
+	/* large_rx disabled for now, additional features needed */
 	adapter->ip_offload_ctrl.large_rx_ipv4 = 0;
 	adapter->ip_offload_ctrl.large_rx_ipv6 = 0;
 
@@ -2648,6 +2679,11 @@ static void handle_query_ip_offload_rsp(struct ibmvnic_adapter *adapter)
 	    (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)))
 		adapter->netdev->features |= NETIF_F_RXCSUM;
 
+	if (buf->large_tx_ipv4)
+		adapter->netdev->hw_features |= NETIF_F_TSO;
+	if (buf->large_tx_ipv6)
+		adapter->netdev->hw_features |= NETIF_F_TSO6;
+
 	memset(&crq, 0, sizeof(crq));
 	crq.control_ip_offload.first = IBMVNIC_CRQ_CMD;
 	crq.control_ip_offload.cmd = CONTROL_IP_OFFLOAD;
-- 
1.8.3.1

Re: [PATCH net-next v3] ibmvnic: Enable TSO support

[David Miller]

From: Thomas Falcon <tlfalcon@linux.vnet.ibm.com>
Date: Fri, 26 May 2017 11:15:46 -0500

> Enable TSO in the ibmvnic driver. Scatter-gather is also enabled,
> though there currently is no support for scatter-gather in
> vNIC-compatible hardware, resulting in forced linearization
> of all fragmented SKB's. Though not ideal, given the throughput
> improvement gained by enabling TSO, it has been decided
> that this is an acceptable tradeoff.
> 
> The feature is also enabled by a module parameter.
> This parameter is necessary because TSO can not easily be
> enabled or disabled in firmware without reinitializing the driver.
> 
> CC: Nathan Fontenot <nfont@linux.vnet.ibm.com>
> CC: John Allen <jallen@linux.vnet.ibm.com>
> Signed-off-by: Thomas Falcon <tlfalcon@linux.vnet.ibm.com>
> ---
> v3: Remove module parameter, TSO will be set through ethtool
>     include ibmvnic_fix_features to manage scatter-gather
> v2: Fix kbuild robot warning from module parameter init

You're only dealing with half of my feedback.

I stated, explicitly, that you're only probably getting a performance
increase with TSO even though it is being software split up before
being sent by the device, because you don't implement xmit_more.

It makes zero sense, to have the networking stack do such an enormous
amount of wasted work putting the TSO segments together just to then
split them apart and linearize the whole packet before transmit.

That should have been a huge red flag for you, and caused you to
inspect further what the true reason is why performance still
increases with all of that happening.

Please look into this further, because I am quite convinced that
you have TX doorbell and VM exit overhead issues at hand, which
would be more properly and suitable dealt with by xmit_more which
is designed _specifically_ to target that apsect of virtual
networking driver performance.

Thanks.