e1000e: adjust PM QoS request
authorBruce Allan <bruce.w.allan@intel.com>
Sat, 12 Jan 2013 07:25:22 +0000 (07:25 +0000)
committerJeff Kirsher <jeffrey.t.kirsher@intel.com>
Mon, 28 Jan 2013 07:27:41 +0000 (23:27 -0800)
It has been found that devices other than 82579 (a.k.a. e1000_pch2lan)
suffer from dropped transactions on platforms with deep C-states when
jumbo frames are enabled.  For example, LOMs on ICH9- and ICH10-based
platforms which recently had early-receive de-featured (for stability
reasons) suffer from this.  To resolve this for all devices, when jumbo
frames are enabled set the PM QoS DMA latency request based on the size
of the receive packet buffer less one full frame.

Signed-off-by: Bruce Allan <bruce.w.allan@intel.com>
Tested-by: Aaron Brown <aaron.f.brown@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
drivers/net/ethernet/intel/e1000e/defines.h
drivers/net/ethernet/intel/e1000e/netdev.c

index 3b99719..c458767 100644 (file)
 #define E1000_PBA_8K  0x0008    /* 8KB */
 #define E1000_PBA_16K 0x0010    /* 16KB */
 
+#define E1000_PBA_RXA_MASK     0xFFFF
+
 #define E1000_PBS_16K E1000_PBA_16K
 
 /* Uncorrectable/correctable ECC Error counts and enable bits */
index e386e95..46a38a4 100644 (file)
@@ -3242,18 +3242,23 @@ static void e1000_configure_rx(struct e1000_adapter *adapter)
                rxcsum &= ~E1000_RXCSUM_TUOFL;
        ew32(RXCSUM, rxcsum);
 
-       if (adapter->hw.mac.type == e1000_pch2lan) {
-               /* With jumbo frames, excessive C-state transition
-                * latencies result in dropped transactions.
-                */
-               if (adapter->netdev->mtu > ETH_DATA_LEN) {
+       /* With jumbo frames, excessive C-state transition latencies result
+        * in dropped transactions.
+        */
+       if (adapter->netdev->mtu > ETH_DATA_LEN) {
+               u32 lat =
+                   ((er32(PBA) & E1000_PBA_RXA_MASK) * 1024 -
+                    adapter->max_frame_size) * 8 / 1000;
+
+               if (adapter->flags & FLAG_IS_ICH) {
                        u32 rxdctl = er32(RXDCTL(0));
                        ew32(RXDCTL(0), rxdctl | 0x3);
-                       pm_qos_update_request(&adapter->netdev->pm_qos_req, 55);
-               } else {
-                       pm_qos_update_request(&adapter->netdev->pm_qos_req,
-                                             PM_QOS_DEFAULT_VALUE);
                }
+
+               pm_qos_update_request(&adapter->netdev->pm_qos_req, lat);
+       } else {
+               pm_qos_update_request(&adapter->netdev->pm_qos_req,
+                                     PM_QOS_DEFAULT_VALUE);
        }
 
        /* Enable Receives */
@@ -4281,10 +4286,8 @@ static int e1000_open(struct net_device *netdev)
                e1000_update_mng_vlan(adapter);
 
        /* DMA latency requirement to workaround jumbo issue */
-       if (adapter->hw.mac.type == e1000_pch2lan)
-               pm_qos_add_request(&adapter->netdev->pm_qos_req,
-                                  PM_QOS_CPU_DMA_LATENCY,
-                                  PM_QOS_DEFAULT_VALUE);
+       pm_qos_add_request(&adapter->netdev->pm_qos_req, PM_QOS_CPU_DMA_LATENCY,
+                          PM_QOS_DEFAULT_VALUE);
 
        /* before we allocate an interrupt, we must be ready to handle it.
         * Setting DEBUG_SHIRQ in the kernel makes it fire an interrupt
@@ -4392,8 +4395,7 @@ static int e1000_close(struct net_device *netdev)
            !test_bit(__E1000_TESTING, &adapter->state))
                e1000e_release_hw_control(adapter);
 
-       if (adapter->hw.mac.type == e1000_pch2lan)
-               pm_qos_remove_request(&adapter->netdev->pm_qos_req);
+       pm_qos_remove_request(&adapter->netdev->pm_qos_req);
 
        pm_runtime_put_sync(&pdev->dev);