dwc_eqos - segmentation offload and vlan fixes (#39)

- Add a README.md.
- Add a REG file for enabling a diagnostic ETW autologger.
- Implement TCP segmentation offload (LSOv2).
- Implement UDP fragmentation offload. (USO).
- Instead of making a half-baked effort to adapt to disabled hardware
  features (i.e. lack of Checksum or Segment offload support), check for
  the feature and don't load if it's missing. At present, I have nothing
  to test this on and it just makes the code messy. We can add
  adaptability back in the future if anybody actually wants it and has
  hardware to test it.
- Fix problems with error logging in Tx path.
- Fix problems with VLAN tag insertion.
- Fix problems with checksum offload configuration. NetAdapterCx only
  recognizes the "granular" options.

drivers/net/dwc_eqos/AutoLogger.reg

@@ -0,0 +1,12 @@
+Windows Registry Editor Version 5.00
+
+[HKEY_LOCAL_MACHINE\SYSTEM\ControlSet001\Control\WMI\Autologger\dwc_eqos]
+"Guid"="{3fde989c-5470-4452-8f3c-91b0584f5a75}"
+"BufferSize"=dword:00000010
+"LogFileMode"=dword:08080400
+"Start"=dword:00000001
+
+[HKEY_LOCAL_MACHINE\SYSTEM\ControlSet001\Control\WMI\Autologger\dwc_eqos\{5d8331d3-70b3-5620-5664-db28f48a4b79}]
+"Enabled"=dword:00000001
+"EnableLevel"=dword:00000004
+"EnableFlags"=dword:00000003

drivers/net/dwc_eqos/README.md

@@ -0,0 +1,48 @@
+# Synopsys DesignWare Ethernet Quality of Service (GMAC) Driver
+
+This is a driver for the Synopsys DesignWare Ethernet Quality of Service (EQoS)
+controller found in the RK35xx SoCs, supporting 1Gbps ethernet connections.
+
+## Compatibility
+
+EQoS is a configurable IP block that can be customized and added to a SoC. This
+driver has been tested only on the RK3588(s) and assumes the presence of
+optional features that may be missing on other SoCs. With minor fixes, it would
+probably work on other EQoS-based SoCs. The driver specifically checks for the
+following:
+
+- `GMAC_MAC_Version.RKVER` must be 0x51 or 0x52 (other values untested).
+- `GMAC_MAC_HW_Feature0.SAVLANINS` must be enabled (require VLAN insertion support).
+- `GMAC_MAC_HW_Feature0.RXCOESEL` and `TXCOESEL` must be enabled (require checksum offload support).
+- `GMAC_MAC_HW_Feature1.TSOEN` must be enabled (require TCP/UDP segmentation offload support).
+
+There may be other requirements that are assumed but not checked.
+
+## ACPI Configuration
+
+This driver uses ACPI properties to configure the EQoS controller's DMA behavior:
+
+- `_DSD\snps,pblx8` (default = 1): Controls the value of `GMAC_DMA_CHx_Control.PBLx8`, i.e. controls whether PBL values are treated as 1-beat units (0) or 8-beat units (1, default).
+- `_DSD\snps,pbl` (default = 8): Default value for `txpbl` and `rxpbl`.
+- `_DSD\snps,txpbl` (default = `pbl`): Controls the value of `GMAC_DMA_CHx_Tx_Control.TxPBL`, i.e. transmit programmable burst length.
+- `_DSD\snps,rxpbl` (default = `pbl`): Controls the value of `GMAC_DMA_CHx_Rx_Control.RxPBL`, i.e. receive programmable burst length.
+- `_DSD\snps,fixed-burst` (default = 0): Controls the value of `GMAC_DMA_SysBus_Mode.FB`.
+- `_DSD\snps,mixed-burst` (default = 1): Controls the value of `GMAC_DMA_SysBus_Mode.Bit14`.
+- `_DSD\snps,axi-config` (default = none): Controls the `$(AXIC)` method name to use for the remaining properties. If not present, the driver will use default values for the remaining properties. Should generally be set to string `"AXIC"`.
+- `$(AXIC)\snps,wr_osr_lmt` (default = 4): Controls the value of `GMAC_DMA_SysBus_Mode.WR_OSR_LMT`, i.e. AXI maximum write outstanding request limit.
+- `$(AXIC)\snps,rd_osr_lmt` (default = 8): Controls the value of `GMAC_DMA_SysBus_Mode.RD_OSR_LMT`, i.e. AXI maximum read outstanding request limit.
+- `$(AXIC)\snps,blen` (default = `{ 16, 8, 4 }`): Controls the values of `GMAC_DMA_SysBus_Mode.BLENx` (x = 4, 8, 16, 32, 64, 128, 256), i.e. AXI burst length. Should be a list of 7 integers, e.g. `Package () { 0, 0, 0, 0, 16, 8, 4 }`.
+
+## Areas for improvement:
+
+- Run against network test suites and fix any issues.
+- Memory optimizations? Current implementation uses system-managed buffers.
+  System-managed buffer size is tied to MTU. When jumbo frames are enabled,
+  this is wasteful since most packets are still 1522 bytes or less. If we
+  used driver-managed buffers and updated the Rx queue to handle multi-buffer
+  packets, we could use 1536-byte or 2048-byte buffers for the Rx queue, saving
+  about 2MB per device when JumboPacket = 9014.
+- Configure speed, duplex via Ndi\params?
+- Power control, wake-on-LAN, ARP offload?
+- Multi-queue, RSS support?
+- Make it more generic (test with other EQoS-based SoCs)?

drivers/net/dwc_eqos/descriptors.h

@@ -21,6 +21,14 @@ enum TxChecksumInsertion : UINT16
     TxChecksumInsertionEnabledIncludingPseudo = 3,
 };
 
+enum TxVlanTagControl : UINT16
+{
+    TxVlanTagControlNone = 0,
+    TxVlanTagControlRemove = 1,
+    TxVlanTagControlInsert = 2,
+    TxVlanTagControlReplace = 3,
+};
+
 struct TxDescriptorRead
 {
     // TDES0, TDES1
@@ -31,7 +39,7 @@ struct TxDescriptorRead
     // TDES2
 
     UINT16 Buf1Length : 14; // B1L
-    UINT16 VlanTagControl : 2; // VTIR
+    TxVlanTagControl VlanTagControl : 2; // VTIR
 
     UINT16 Buf2Length : 14; // B2L
     UINT16 TransmitTimestampEnable : 1; // TTSE
@@ -68,7 +76,7 @@ struct TxDescriptorReadTso
     // TDES2
 
     UINT16 Buf1Length : 14; // B1L (10-bit header length if FD = 1)
-    UINT16 VlanTagControl : 2; // VTIR
+    TxVlanTagControl VlanTagControl : 2; // VTIR
 
     UINT16 Buf2Length : 14; // B2L
     UINT16 TsoMemoryWriteDisable : 1; // TMWD
@@ -77,7 +85,7 @@ struct TxDescriptorReadTso
     // TDES3
 
     UINT32 TcpPayloadLength : 18; // TPL
-    UINT32 TcpSegmentationEnable : 1; // TSE = 0
+    UINT32 TcpSegmentationEnable : 1; // TSE = 1
     UINT32 TcpHeaderLength : 4; // TCP/UDP header length (must be 2 for UDP)
     UINT32 SourceAddressInsertionControl : 3; // SAIC
     UINT32 Reserved26 : 2; // CPC, ignored when TSE = 1
@@ -159,7 +167,7 @@ struct TxDescriptorContext
 
     UINT8 VlanTagValid : 1; // VLTV
     UINT8 InnerVlanTagValid : 1; // IVLTV
-    UINT8 InnverVlanTagControl : 2; // IVTIR
+    UINT8 InnerVlanTagControl : 2; // IVTIR
     UINT8 Reserved20 : 3;
     UINT8 DescriptorError : 1; // DE
 

drivers/net/dwc_eqos/device.cpp

@@ -574,6 +574,25 @@ AdapterOffloadSetRxChecksum(
         TraceLoggingBoolean(Udp));
 }
 
+static EVT_NET_ADAPTER_OFFLOAD_SET_GSO AdapterOffloadSetGso;
+static void
+AdapterOffloadSetGso(
+    _In_ NETADAPTER adapter,
+    _In_ NETOFFLOAD offload)
+{
+    // PASSIVE_LEVEL, nonpaged (resume path)
+    UNREFERENCED_PARAMETER(adapter);
+    auto const LsoIPv4 = NetOffloadIsLsoIPv4Enabled(offload);
+    auto const LsoIPv6 = NetOffloadIsLsoIPv6Enabled(offload);
+    auto const UsoIPv4 = NetOffloadIsUsoIPv4Enabled(offload);
+    auto const UsoIPv6 = NetOffloadIsUsoIPv6Enabled(offload);
+    TraceEntryExit(AdapterOffloadSetGso, LEVEL_INFO,
+        TraceLoggingBoolean(LsoIPv4),
+        TraceLoggingBoolean(LsoIPv6),
+        TraceLoggingBoolean(UsoIPv4),
+        TraceLoggingBoolean(UsoIPv6));
+}
+
 static EVT_WDF_DEVICE_D0_ENTRY DeviceD0Entry;
 static NTSTATUS
 DeviceD0Entry(
@@ -644,7 +663,7 @@ DeviceD0Entry(
     //macConfig.PadOrCrcStripEnable = true;   // Why doesn't this work?
     //macConfig.CrcStripEnableForType = true; // Why doesn't this work?
     macConfig.GiantPacketSizeLimitControlEnable = context->config.jumboFrame > JumboPacketMin;
-    macConfig.ChecksumOffloadEnable = context->config.txCoeSel || context->config.rxCoeSel;
+    macConfig.ChecksumOffloadEnable = true;
     Write32(&context->regs->Mac_Configuration, macConfig);
 
     MacExtConfiguration_t macExtConfig = {};
@@ -960,13 +979,35 @@ DevicePrepareHardware(
             status = STATUS_DEVICE_CONFIGURATION_ERROR;
             goto Done;
         }
+
+        if (!context->feature0.TxChecksumOffload)
+        {
+            // Could adapt at runtime if needed, but assume it's present for now.
+            TraceWrite("DevicePrepareHardware-TxChecksumOffload-required", LEVEL_ERROR);
+            status = STATUS_DEVICE_CONFIGURATION_ERROR;
+            goto Done;
+        }
+
+        if (!context->feature0.RxChecksumOffload)
+        {
+            // Could adapt at runtime if needed, but assume it's present for now.
+            TraceWrite("DevicePrepareHardware-RxChecksumOffload-required", LEVEL_ERROR);
+            status = STATUS_DEVICE_CONFIGURATION_ERROR;
+            goto Done;
+        }
+
+        if (!context->feature1.TsoEn)
+        {
+            // Could adapt at runtime if needed, but assume it's present for now.
+            TraceWrite("DevicePrepareHardware-TsoEn-required", LEVEL_ERROR);
+            status = STATUS_DEVICE_CONFIGURATION_ERROR;
+            goto Done;
+        }
     }
 
     // Device Config
 
     {
-        context->config.txCoeSel = context->feature0.TxChecksumOffload;
-        context->config.rxCoeSel = context->feature0.RxChecksumOffload;
         context->config.pblX8 = true;
         context->config.pbl = 8;
         context->config.txPbl = context->config.pbl;
@@ -1222,7 +1263,7 @@ DevicePrepareHardware(
 
         NET_ADAPTER_TX_CAPABILITIES txCaps;
         NET_ADAPTER_TX_CAPABILITIES_INIT_FOR_DMA(&txCaps, &dmaCaps, QueuesSupported);
-        txCaps.MaximumNumberOfFragments = QueueDescriptorMinCount - 2; // = 1 hole in the ring + 1 context descriptor.
+        txCaps.MaximumNumberOfFragments = TxMaximumNumberOfFragments;
 
         // TODO: Driver-managed buffering + multi-descriptor receive would
         // reduce memory overhead of Jumbo Packets.
@@ -1246,21 +1287,18 @@ DevicePrepareHardware(
             NetPacketFilterFlagPromiscuous;
         NetAdapterSetReceiveFilterCapabilities(context->adapter, &rxFilterCaps);
 
-        if (context->config.txCoeSel)
-        {
-            NET_ADAPTER_OFFLOAD_TX_CHECKSUM_CAPABILITIES txChecksumCaps;
-            NET_ADAPTER_OFFLOAD_TX_CHECKSUM_CAPABILITIES_INIT(&txChecksumCaps, {}, AdapterOffloadSetTxChecksum);
-            txChecksumCaps.Layer3Flags =
-                NetAdapterOffloadLayer3FlagIPv4NoOptions |
-                NetAdapterOffloadLayer3FlagIPv4WithOptions |
-                NetAdapterOffloadLayer3FlagIPv6NoExtensions |
-                NetAdapterOffloadLayer3FlagIPv6WithExtensions;
-            txChecksumCaps.Layer4Flags =
-                NetAdapterOffloadLayer4FlagTcpNoOptions |
-                NetAdapterOffloadLayer4FlagTcpWithOptions |
-                NetAdapterOffloadLayer4FlagUdp;
-            NetAdapterOffloadSetTxChecksumCapabilities(context->adapter, &txChecksumCaps);
-        }
+        NET_ADAPTER_OFFLOAD_TX_CHECKSUM_CAPABILITIES txChecksumCaps;
+        NET_ADAPTER_OFFLOAD_TX_CHECKSUM_CAPABILITIES_INIT(&txChecksumCaps,
+            NetAdapterOffloadLayer3FlagIPv4NoOptions |
+            NetAdapterOffloadLayer3FlagIPv4WithOptions |
+            NetAdapterOffloadLayer3FlagIPv6NoExtensions |
+            NetAdapterOffloadLayer3FlagIPv6WithExtensions,
+            AdapterOffloadSetTxChecksum);
+        txChecksumCaps.Layer4Flags =
+            NetAdapterOffloadLayer4FlagTcpNoOptions |
+            NetAdapterOffloadLayer4FlagTcpWithOptions |
+            NetAdapterOffloadLayer4FlagUdp;
+        NetAdapterOffloadSetTxChecksumCapabilities(context->adapter, &txChecksumCaps);
 
         NET_ADAPTER_OFFLOAD_RX_CHECKSUM_CAPABILITIES rxChecksumCaps;
         NET_ADAPTER_OFFLOAD_RX_CHECKSUM_CAPABILITIES_INIT(&rxChecksumCaps,
@@ -1272,6 +1310,21 @@ DevicePrepareHardware(
             NetAdapterOffloadIeee8021PriorityTaggingFlag |
             NetAdapterOffloadIeee8021VlanTaggingFlag);
         NetAdapterOffloadSetIeee8021qTagCapabilities(context->adapter, &ieee8021qCaps);
+
+        NET_ADAPTER_OFFLOAD_GSO_CAPABILITIES gsoCaps;
+        NET_ADAPTER_OFFLOAD_GSO_CAPABILITIES_INIT(&gsoCaps,
+            NetAdapterOffloadLayer3FlagIPv4NoOptions |
+            NetAdapterOffloadLayer3FlagIPv4WithOptions |
+            NetAdapterOffloadLayer3FlagIPv6NoExtensions |
+            NetAdapterOffloadLayer3FlagIPv6WithExtensions,
+            NetAdapterOffloadLayer4FlagUdp |
+            NetAdapterOffloadLayer4FlagTcpNoOptions |
+            NetAdapterOffloadLayer4FlagTcpWithOptions,
+            TxMaximumOffloadSize,
+            2, // MinimumSegmentCount
+            AdapterOffloadSetGso);
+        gsoCaps.Layer4HeaderOffsetLimit = TxLayer4HeaderOffsetLimit;
+        NetAdapterOffloadSetGsoCapabilities(context->adapter, &gsoCaps);
     }
 
     // Initialize adapter.

drivers/net/dwc_eqos/device.h

@@ -8,8 +8,6 @@ struct DeviceContext; // TODO: if we do multi-queue, make a DeviceQueueContext s
 // Information about the device provided to the queues.
 struct DeviceConfig
 {
-    bool txCoeSel;      // MAC_HW_Feature0\TXCOESEL (hardware support for tx checksum offload).
-    bool rxCoeSel;      // MAC_HW_Feature0\RXCOESEL (hardware support for rx checksum offload).
     bool pblX8;         // _DSD\snps,pblx8 (default = 1).
     UINT8 pbl;          // _DSD\snps,pbl (default = 8; effect depends on pblX8).
     UINT8 txPbl;        // _DSD\snps,txpbl (default = pbl; effect depends on pblX8).
@@ -21,7 +19,7 @@ struct DeviceConfig
     UINT8 blen : 7;     // AXIC\snps,blen bitmask of 7 booleans 4..256 (default = 4, 8, 16).
     bool txFlowControl; // Adapter configuration (Ndi\params\*FlowControl).
     bool rxFlowControl; // Adapter configuration (Ndi\params\*FlowControl).
-    UINT16 jumboFrame;  // Adapter configuration (Ndi\params\*JumboFrame). 1514..4088
+    UINT16 jumboFrame;  // Adapter configuration (Ndi\params\*JumboFrame). 1514..9014.
 
     UINT16 RxBufferSize() const
     {

drivers/net/dwc_eqos/driver.cpp

@@ -2,15 +2,6 @@
 #include "device.h"
 #include "trace.h"
 
-/*
-Possible areas for improvement:
-- Tx segmentation offload.
-- Run against network test suites and fix any issues.
-- Power control, wake-on-LAN, ARP offload.
-- Configure speed, duplex in Ndi\params.
-- Multi-queue support?
-*/
-
 TRACELOGGING_DEFINE_PROVIDER(
     TraceProvider,
     "dwc_eqos",