Add support for setting TypeOfService/TrafficClass on connections

docs/Settings.md

@@ -177,6 +177,7 @@ These parameters are accessed by calling [GetParam](./api/GetParam.md) or [SetPa
 | `QUIC_PARAM_CONN_STATISTICS_V2`<br> 22            | QUIC_STATISTICS_V2            | Get-only  | Connection-level statistics, version 2.                                                   |
 | `QUIC_PARAM_CONN_STATISTICS_V2_PLAT`<br> 23       | QUIC_STATISTICS_V2            | Get-only  | Connection-level statistics with platform-specific time format, version 2.                |
 | `QUIC_PARAM_CONN_ORIG_DEST_CID` <br> 24           | uint8_t[]                     | Get-only  | The original destination connection ID used by the client to connect to the server.       |
+| `QUIC_PARAM_CONN_TYPE_OF_SERVICE` <br> 25         | uint8_t                       | Both      | The value put in the Type of Service/Traffic Class field on packets sent from this connection. |
 
 ### QUIC_PARAM_CONN_STATISTICS_V2
 

src/core/connection.c

@@ -1837,6 +1837,7 @@ QuicConnStart(
     UdpConfig.Flags = Connection->State.ShareBinding ? CXPLAT_SOCKET_FLAG_SHARE : 0;
     UdpConfig.InterfaceIndex = Connection->State.LocalInterfaceSet ? (uint32_t)Path->Route.LocalAddress.Ipv6.sin6_scope_id : 0, // NOLINT(google-readability-casting)
     UdpConfig.PartitionIndex = QuicPartitionIdGetIndex(Connection->PartitionID);
+    UdpConfig.TypeOfService = Connection->TypeOfService;
 #ifdef QUIC_COMPARTMENT_ID
     UdpConfig.CompartmentId = Configuration->CompartmentId;
 #endif
@@ -6218,6 +6219,7 @@ QuicConnParamSet(
             UdpConfig.RemoteAddress = &Connection->Paths[0].Route.RemoteAddress;
             UdpConfig.Flags = Connection->State.ShareBinding ? CXPLAT_SOCKET_FLAG_SHARE : 0;
             UdpConfig.InterfaceIndex = 0;
+            UdpConfig.TypeOfService = Connection->TypeOfService;
 #ifdef QUIC_COMPARTMENT_ID
             UdpConfig.CompartmentId = Connection->Configuration->CompartmentId;
 #endif
@@ -6623,6 +6625,34 @@ QuicConnParamSet(
         return QUIC_STATUS_SUCCESS;
     }
 
+    case QUIC_PARAM_CONN_TYPE_OF_SERVICE:
+    {
+        if (BufferLength != sizeof(uint8_t) || Buffer == NULL) {
+            Status = QUIC_STATUS_INVALID_PARAMETER;
+            break;
+        }
+
+        uint8_t TypeOfService = 0;
+        CxPlatCopyMemory(&TypeOfService, Buffer, BufferLength);
+
+        if (TypeOfService > CXPLAT_MAX_TYPE_OF_SERVICE) {
+            Status = QUIC_STATUS_INVALID_PARAMETER;
+            break;
+        }
+
+        Connection->TypeOfService = TypeOfService;
+
+        if (Connection->State.Started) {
+            Status =
+                CxPlatSocketSetTypeOfService(
+                    Connection->Paths[0].Binding->Socket,
+                    TypeOfService);
+        } else {
+            Status = QUIC_STATUS_SUCCESS;
+        }
+        break;
+    }
+
     //
     // Private
     //
@@ -7207,27 +7237,55 @@ QuicConnParamGet(
     }
 
     case QUIC_PARAM_CONN_ORIG_DEST_CID:
+
         if (Connection->OrigDestCID == NULL) {
             Status = QUIC_STATUS_INVALID_STATE;
             break;
         }
+
         if (*BufferLength < Connection->OrigDestCID->Length) {
             Status = QUIC_STATUS_BUFFER_TOO_SMALL;
             *BufferLength = Connection->OrigDestCID->Length;
             break;
         }
+
         if (Buffer == NULL) {
             Status = QUIC_STATUS_INVALID_PARAMETER;
             break;
         }
+
         CxPlatCopyMemory(
             Buffer,
             Connection->OrigDestCID->Data,
             Connection->OrigDestCID->Length);
+
         //
         // Tell app how much buffer we copied.
         //
         *BufferLength = Connection->OrigDestCID->Length;
+
+        Status = QUIC_STATUS_SUCCESS;
+        break;
+
+     case QUIC_PARAM_CONN_TYPE_OF_SERVICE:
+
+        if (*BufferLength < sizeof(Connection->TypeOfService)) {
+            Status = QUIC_STATUS_BUFFER_TOO_SMALL;
+            *BufferLength = sizeof(Connection->TypeOfService);
+            break;
+        }
+
+        if (Buffer == NULL) {
+            Status = QUIC_STATUS_INVALID_PARAMETER;
+            break;
+        }
+
+        CxPlatCopyMemory(
+            Buffer,
+            &Connection->TypeOfService,
+            sizeof(Connection->TypeOfService));
+
+        *BufferLength = sizeof(Connection->TypeOfService);
         Status = QUIC_STATUS_SUCCESS;
         break;
 

src/core/connection.h

@@ -462,6 +462,12 @@ typedef struct QUIC_CONNECTION {
     //
     uint8_t PeerReorderingThreshold;
 
+    //
+    // Type of Service value to set on the socket when the connection is started.
+    // Default value of 0.
+    //
+    uint8_t TypeOfService;
+
     //
     // The ACK frequency sequence number we are currently using to send.
     //

src/inc/msquic.h

@@ -923,6 +923,7 @@ typedef struct QUIC_SCHANNEL_CREDENTIAL_ATTRIBUTE_W {
 #define QUIC_PARAM_CONN_STATISTICS_V2                   0x05000016  // QUIC_STATISTICS_V2
 #define QUIC_PARAM_CONN_STATISTICS_V2_PLAT              0x05000017  // QUIC_STATISTICS_V2
 #define QUIC_PARAM_CONN_ORIG_DEST_CID                   0x05000018  // uint8_t[]
+#define QUIC_PARAM_CONN_TYPE_OF_SERVICE                 0x50000019  // uint8_t
 
 //
 // Parameters for TLS.

src/inc/quic_datapath.h

@@ -58,6 +58,12 @@ typedef enum CXPLAT_ECN_TYPE {
 //
 #define CXPLAT_ECN_FROM_TOS(ToS) (CXPLAT_ECN_TYPE)((ToS) & 0x3)
 
+//
+// Define the maximum type of service value allowed.
+// Note: this is without the ECN bits included
+//
+#define CXPLAT_MAX_TYPE_OF_SERVICE 63
+
 //
 // The maximum IP MTU this implementation supports for QUIC.
 //
@@ -444,6 +450,7 @@ CxPlatDataPathUpdateConfig(
 #define CXPLAT_DATAPATH_FEATURE_TCP                   0x0020
 #define CXPLAT_DATAPATH_FEATURE_RAW                   0x0040
 #define CXPLAT_DATAPATH_FEATURE_TTL                   0x0080
+#define CXPLAT_DATAPATH_FEATURE_TYPE_OF_SERVICE       0x0100
 
 //
 // Queries the currently supported features of the datapath.
@@ -544,6 +551,7 @@ typedef struct CXPLAT_UDP_CONFIG {
 #ifdef QUIC_OWNING_PROCESS
     QUIC_PROCESS OwningProcess;         // Kernel client-only
 #endif
+    uint8_t TypeOfService;              // Default 0. Optional.
 
     // used for RAW datapath
     uint8_t CibirIdLength;              // CIBIR ID length. Value of 0 indicates CIBIR isn't used
@@ -646,6 +654,17 @@ CxPlatSocketGetRemoteAddress(
     _Out_ QUIC_ADDR* Address
     );
 
+//
+// Sets TypeOfService on the socket. May fail if insufficient privileges exist
+// to set the desired value on the socket.
+//
+_IRQL_requires_max_(PASSIVE_LEVEL)
+QUIC_STATUS
+CxPlatSocketSetTypeOfService(
+    _In_ CXPLAT_SOCKET* Socket,
+    _In_ uint8_t TypeOfService
+    );
+
 //
 // Queries a raw socket availability.
 //

src/platform/datapath_epoll.c

@@ -340,6 +340,7 @@ CxPlatDataPathCalculateFeatureSupport(
 
     Datapath->Features |= CXPLAT_DATAPATH_FEATURE_TCP;
     Datapath->Features |= CXPLAT_DATAPATH_FEATURE_TTL;
+    Datapath->Features |= CXPLAT_DATAPATH_FEATURE_TYPE_OF_SERVICE;
 }
 
 void
@@ -887,6 +888,45 @@ CxPlatSocketContextInitialize(
             goto Exit;
         }
 
+        if (Config->TypeOfService != 0) {
+            Option = Config->TypeOfService;
+            Result =
+                setsockopt(
+                    SocketContext->SocketFd,
+                    IPPROTO_IP,
+                    IP_TOS,
+                    (const void*)&Option,
+                    sizeof(Option));
+            if (Result == SOCKET_ERROR) {
+                Status = errno;
+                QuicTraceEvent(
+                    DatapathErrorStatus,
+                    "[data][%p] ERROR, %u, %s.",
+                    Binding,
+                    Status,
+                    "setsockopt(IP_TOS) failed");
+                goto Exit;
+            }
+
+            Option = Config->TypeOfService;
+            Result =
+                setsockopt(
+                    SocketContext->SocketFd,
+                    IPPROTO_IPV6,
+                    IPV6_TCLASS,
+                    (const void*)&Option,
+                    sizeof(Option));
+            if (Result == SOCKET_ERROR) {
+                Status = errno;
+                QuicTraceEvent(
+                    DatapathErrorStatus,
+                    "[data][%p] ERROR, %u, %s.",
+                    Binding,
+                    Status,
+                    "setsockopt(IPV6_TCLASS) failed");
+                goto Exit;
+            }
+        }
 
     #ifdef UDP_GRO
         if (SocketContext->DatapathPartition->Datapath->Features & CXPLAT_DATAPATH_FEATURE_RECV_COALESCING) {
@@ -2455,6 +2495,60 @@ SocketSend(
     }
 }
 
+_IRQL_requires_max_(PASSIVE_LEVEL)
+QUIC_STATUS
+SocketSetTypeOfService(
+    _In_ CXPLAT_SOCKET* Socket,
+    _In_ uint8_t TypeOfService
+    )
+{
+    const uint16_t SocketCount =  Socket->NumPerProcessorSockets ? (uint16_t)CxPlatProcCount() : 1;
+    QUIC_STATUS Status = QUIC_STATUS_SUCCESS;
+    for (uint16_t i = 0; i < SocketCount; i++) {
+        CXPLAT_SOCKET_CONTEXT* SocketContext = &Socket->SocketContexts[i];
+        int Option = TypeOfService;
+        int if (Config->TypeOfService != 0) {
+            Option = Config->TypeOfService;
+            Result =
+                setsockopt(
+                    SocketContext->SocketFd,
+                    IPPROTO_IP,
+                    IP_TOS,
+                    (const void*)&Option,
+                    sizeof(Option));
+            if (Result == SOCKET_ERROR) {
+                Status = errno;
+                QuicTraceEvent(
+                    DatapathErrorStatus,
+                    "[data][%p] ERROR, %u, %s.",
+                    Binding,
+                    Status,
+                    "setsockopt(IP_TOS) failed");
+                break;
+            }
+
+            Option = TypeOfService;
+            Result =
+                setsockopt(
+                    SocketContext->SocketFd,
+                    IPPROTO_IPV6,
+                    IPV6_TCLASS,
+                    (const void*)&Option,
+                    sizeof(Option));
+            if (Result == SOCKET_ERROR) {
+                Status = errno;
+                QuicTraceEvent(
+                    DatapathErrorStatus,
+                    "[data][%p] ERROR, %u, %s.",
+                    Binding,
+                    Status,
+                    "setsockopt(IPV6_TCLASS) failed");
+                break;
+            }
+        }
+    }
+    return Status;
+}
 //
 // This is defined and used instead of CMSG_NXTHDR because (1) we've already
 // done the work to ensure the necessary space is available and (2) CMSG_NXTHDR

src/platform/datapath_raw.c

@@ -371,7 +371,7 @@ RawSocketSend(
     const CXPLAT_INTERFACE* Interface = CxPlatDpRawGetInterfaceFromQueue(Route->Queue);
 
     CxPlatFramingWriteHeaders(
-        Socket, Route, &SendData->Buffer, SendData->ECN,
+        Socket, Route, &SendData->Buffer, SendData->ECN, Socket->TypeOfService,
         Interface->OffloadStatus.Transmit.NetworkLayerXsum,
         Interface->OffloadStatus.Transmit.TransportLayerXsum,
         Route->TcpState.SequenceNumber,
@@ -380,3 +380,15 @@ RawSocketSend(
     CxPlatDpRawTxEnqueue(SendData);
     return QUIC_STATUS_SUCCESS;
 }
+
+
+_IRQL_requires_max_(DISPATCH_LEVEL)
+QUIC_STATUS
+RawSocketSetTypeOfService(
+    _In_ CXPLAT_SOCKET_RAW* Socket,
+    _In_ uint8_t TypeOfService
+    )
+{
+    Socket->TypeOfService = TypeOfService;
+    return QUIC_STATUS_SUCCESS;
+}

src/platform/datapath_raw.h

@@ -265,6 +265,7 @@ typedef struct CXPLAT_SOCKET_RAW {
     SOCKET AuxSocket;
     BOOLEAN Wildcard;                // Using a wildcard local address. Optimization
                                      // to avoid always reading LocalAddress.
+    uint8_t TypeOfService;           // IP TypeOfService/IPv6 Traffic Class value. Default 0.
     uint8_t CibirIdLength;           // CIBIR ID length. Value of 0 indicates CIBIR isn't used
     uint8_t CibirIdOffsetSrc;        // CIBIR ID offset in source CID
     uint8_t CibirIdOffsetDst;        // CIBIR ID offset in destination CID
@@ -377,6 +378,7 @@ CxPlatFramingWriteHeaders(
     _In_ const CXPLAT_ROUTE* Route,
     _Inout_ QUIC_BUFFER* Buffer,
     _In_ CXPLAT_ECN_TYPE ECN,
+    _In_ uint8_t TypeOfService,
     _In_ BOOLEAN SkipNetworkLayerXsum,
     _In_ BOOLEAN SkipTransportLayerXsum,
     _In_ uint32_t TcpSeqNum,

src/platform/datapath_raw_linux.c

@@ -117,6 +117,7 @@ RawSocketCreateUdp(
     NewSocket->CibirIdLength = Config->CibirIdLength;
     NewSocket->CibirIdOffsetSrc = Config->CibirIdOffsetSrc;
     NewSocket->CibirIdOffsetDst = Config->CibirIdOffsetDst;
+    NewSocket->TypeOfService = Config->TypeOfService;
     NewSocket->AuxSocket = INVALID_SOCKET;
     NewSocket->UseTcp = Raw->UseTcp;
     if (Config->CibirIdLength) {