ixgbe: enable extremely low latency

index 7b391f9..b72a997 100644
@@ -2079,6 +2079,27 @@ static int ixgbe_get_coalesce(struct net_device *netdev,
 	return 0;
 }
 
+/*
+ * this function must be called before setting the new value of
+ * rx_itr_setting
+ */
+static void ixgbe_reenable_rsc(struct ixgbe_adapter *adapter,
+                               struct ethtool_coalesce *ec)
+{
+	/* check the old value and enable RSC if necessary */
+	if ((adapter->rx_itr_setting == 0) &&
+	    (adapter->flags2 & IXGBE_FLAG2_RSC_CAPABLE)) {
+		adapter->flags2 |= IXGBE_FLAG2_RSC_ENABLED;
+		adapter->netdev->features |= NETIF_F_LRO;
+		DPRINTK(PROBE, INFO, "rx-usecs set to %d, re-enabling RSC\n",
+		        ec->rx_coalesce_usecs);
+		if (netif_running(adapter->netdev))
+			ixgbe_reinit_locked(adapter);
+		else
+			ixgbe_reset(adapter);
+	}
+}
+
 static int ixgbe_set_coalesce(struct net_device *netdev,
                               struct ethtool_coalesce *ec)
 {

@@ -2095,11 +2116,20 @@ static int ixgbe_set_coalesce(struct net_device *netdev,
 		adapter->tx_ring[0]->work_limit = ec->tx_max_coalesced_frames_irq;
 
 	if (ec->rx_coalesce_usecs > 1) {
+		u32 max_int;
+		if (adapter->flags2 & IXGBE_FLAG2_RSC_ENABLED)
+			max_int = IXGBE_MAX_RSC_INT_RATE;
+		else
+			max_int = IXGBE_MAX_INT_RATE;
+
 		/* check the limits */
-		if ((1000000/ec->rx_coalesce_usecs > IXGBE_MAX_INT_RATE) ||
+		if ((1000000/ec->rx_coalesce_usecs > max_int) ||
 		    (1000000/ec->rx_coalesce_usecs < IXGBE_MIN_INT_RATE))
 			return -EINVAL;
 
+		/* check the old value and enable RSC if necessary */
+		ixgbe_reenable_rsc(adapter, ec);
+
 		/* store the value in ints/second */
 		adapter->rx_eitr_param = 1000000/ec->rx_coalesce_usecs;
 

@@ -2108,6 +2138,9 @@ static int ixgbe_set_coalesce(struct net_device *netdev,
 		/* clear the lower bit as its used for dynamic state */
 		adapter->rx_itr_setting &= ~1;
 	} else if (ec->rx_coalesce_usecs == 1) {
+		/* check the old value and enable RSC if necessary */
+		ixgbe_reenable_rsc(adapter, ec);
+
 		/* 1 means dynamic mode */
 		adapter->rx_eitr_param = 20000;
 		adapter->rx_itr_setting = 1;

@@ -2116,14 +2149,34 @@ static int ixgbe_set_coalesce(struct net_device *netdev,
 		 * any other value means disable eitr, which is best
 		 * served by setting the interrupt rate very high
 		 */
-		if (adapter->flags2 & IXGBE_FLAG2_RSC_ENABLED)
-			adapter->rx_eitr_param = IXGBE_MAX_RSC_INT_RATE;
-		else
-			adapter->rx_eitr_param = IXGBE_MAX_INT_RATE;
+		adapter->rx_eitr_param = IXGBE_MAX_INT_RATE;
 		adapter->rx_itr_setting = 0;
+
+		/*
+		 * if hardware RSC is enabled, disable it when
+		 * setting low latency mode, to avoid errata, assuming
+		 * that when the user set low latency mode they want
+		 * it at the cost of anything else
+		 */
+		if (adapter->flags2 & IXGBE_FLAG2_RSC_ENABLED) {
+			adapter->flags2 &= ~IXGBE_FLAG2_RSC_ENABLED;
+			netdev->features &= ~NETIF_F_LRO;
+			DPRINTK(PROBE, INFO,
+			        "rx-usecs set to 0, disabling RSC\n");
+
+			if (netif_running(netdev))
+				ixgbe_reinit_locked(adapter);
+			else
+				ixgbe_reset(adapter);
+			return 0;
+		}
 	}
 
 	if (ec->tx_coalesce_usecs > 1) {
+		/*
+		 * don't have to worry about max_int as above because
+		 * tx vectors don't do hardware RSC (an rx function)
+		 */
 		/* check the limits */
 		if ((1000000/ec->tx_coalesce_usecs > IXGBE_MAX_INT_RATE) ||
 		    (1000000/ec->tx_coalesce_usecs < IXGBE_MIN_INT_RATE))

@@ -2178,10 +2231,26 @@ static int ixgbe_set_flags(struct net_device *netdev, u32 data)
 	ethtool_op_set_flags(netdev, data);
 
 	/* if state changes we need to update adapter->flags and reset */
-	if ((!!(data & ETH_FLAG_LRO)) != 
-	    (!!(adapter->flags2 & IXGBE_FLAG2_RSC_ENABLED))) {
-		adapter->flags2 ^= IXGBE_FLAG2_RSC_ENABLED;
-		need_reset = true;
+	if (adapter->flags2 & IXGBE_FLAG2_RSC_CAPABLE) {
+		/*
+		 * cast both to bool and verify if they are set the same
+		 * but only enable RSC if itr is non-zero, as
+		 * itr=0 and RSC are mutually exclusive
+		 */
+		if (((!!(data & ETH_FLAG_LRO)) !=
+		     (!!(adapter->flags2 & IXGBE_FLAG2_RSC_ENABLED))) &&
+		    adapter->rx_itr_setting) {
+			adapter->flags2 ^= IXGBE_FLAG2_RSC_ENABLED;
+			switch (adapter->hw.mac.type) {
+			case ixgbe_mac_82599EB:
+				need_reset = true;
+				break;
+			default:
+				break;
+			}
+		} else if (!adapter->rx_itr_setting) {
+			netdev->features &= ~ETH_FLAG_LRO;
+		}
 	}
 
 	/*

index 32781b3..0c4ca68 100644
@@ -1190,6 +1190,15 @@ void ixgbe_write_eitr(struct ixgbe_q_vector *q_vector)
 		itr_reg |= (itr_reg << 16);
 	} else if (adapter->hw.mac.type == ixgbe_mac_82599EB) {
 		/*
+		 * 82599 can support a value of zero, so allow it for
+		 * max interrupt rate, but there is an errata where it can
+		 * not be zero with RSC
+		 */
+		if (itr_reg == 8 &&
+		    !(adapter->flags2 & IXGBE_FLAG2_RSC_ENABLED))
+			itr_reg = 0;
+
+		/*
 		 * set the WDIS bit to not clear the timer bits and cause an
 		 * immediate assertion of the interrupt
 		 */

index b72a997..dfbfe35 100644
@@ -2083,7 +2083,7 @@ static int ixgbe_get_coalesce(struct net_device *netdev,
  * this function must be called before setting the new value of
  * rx_itr_setting
  */
-static void ixgbe_reenable_rsc(struct ixgbe_adapter *adapter,
+static bool ixgbe_reenable_rsc(struct ixgbe_adapter *adapter,
                                struct ethtool_coalesce *ec)
 {
 	/* check the old value and enable RSC if necessary */

@@ -2093,11 +2093,9 @@ static void ixgbe_reenable_rsc(struct ixgbe_adapter *adapter,
 		adapter->netdev->features |= NETIF_F_LRO;
 		DPRINTK(PROBE, INFO, "rx-usecs set to %d, re-enabling RSC\n",
 		        ec->rx_coalesce_usecs);
-		if (netif_running(adapter->netdev))
-			ixgbe_reinit_locked(adapter);
-		else
-			ixgbe_reset(adapter);
+		return true;
 	}
+	return false;
 }
 
 static int ixgbe_set_coalesce(struct net_device *netdev,

@@ -2106,6 +2104,7 @@ static int ixgbe_set_coalesce(struct net_device *netdev,
 	struct ixgbe_adapter *adapter = netdev_priv(netdev);
 	struct ixgbe_q_vector *q_vector;
 	int i;
+	bool need_reset = false;
 
 	/* don't accept tx specific changes if we've got mixed RxTx vectors */
 	if (adapter->q_vector[0]->txr_count && adapter->q_vector[0]->rxr_count

@@ -2128,7 +2127,7 @@ static int ixgbe_set_coalesce(struct net_device *netdev,
 			return -EINVAL;
 
 		/* check the old value and enable RSC if necessary */
-		ixgbe_reenable_rsc(adapter, ec);
+		need_reset = ixgbe_reenable_rsc(adapter, ec);
 
 		/* store the value in ints/second */
 		adapter->rx_eitr_param = 1000000/ec->rx_coalesce_usecs;

@@ -2139,7 +2138,7 @@ static int ixgbe_set_coalesce(struct net_device *netdev,
 		adapter->rx_itr_setting &= ~1;
 	} else if (ec->rx_coalesce_usecs == 1) {
 		/* check the old value and enable RSC if necessary */
-		ixgbe_reenable_rsc(adapter, ec);
+		need_reset = ixgbe_reenable_rsc(adapter, ec);
 
 		/* 1 means dynamic mode */
 		adapter->rx_eitr_param = 20000;

@@ -2164,11 +2163,7 @@ static int ixgbe_set_coalesce(struct net_device *netdev,
 			DPRINTK(PROBE, INFO,
 			        "rx-usecs set to 0, disabling RSC\n");
 
-			if (netif_running(netdev))
-				ixgbe_reinit_locked(adapter);
-			else
-				ixgbe_reset(adapter);
-			return 0;
+			need_reset = true;
 		}
 	}
 

@@ -2220,6 +2215,18 @@ static int ixgbe_set_coalesce(struct net_device *netdev,
 		ixgbe_write_eitr(q_vector);
 	}
 
+	/*
+	 * do reset here at the end to make sure EITR==0 case is handled
+	 * correctly w.r.t stopping tx, and changing TXDCTL.WTHRESH settings
+	 * also locks in RSC enable/disable which requires reset
+	 */
+	if (need_reset) {
+		if (netif_running(netdev))
+			ixgbe_reinit_locked(adapter);
+		else
+			ixgbe_reset(adapter);
+	}
+
 	return 0;
 }
 

index 0c4ca68..fe7e260 100644
@@ -2945,8 +2945,13 @@ static int ixgbe_up_complete(struct ixgbe_adapter *adapter)
 	for (i = 0; i < adapter->num_tx_queues; i++) {
 		j = adapter->tx_ring[i]->reg_idx;
 		txdctl = IXGBE_READ_REG(hw, IXGBE_TXDCTL(j));
-		/* enable WTHRESH=8 descriptors, to encourage burst writeback */
-		txdctl |= (8 << 16);
+		if (adapter->rx_itr_setting == 0) {
+			/* cannot set wthresh when itr==0 */
+			txdctl &= ~0x007F0000;
+		} else {
+			/* enable WTHRESH=8 descriptors, to encourage burst writeback */
+			txdctl |= (8 << 16);
+		}
 		IXGBE_WRITE_REG(hw, IXGBE_TXDCTL(j), txdctl);
 	}