datapath-windows/ovsext/Vxlan.c

   1 /*
   2  * Copyright (c) 2014 VMware, Inc.
   3  *
   4  * Licensed under the Apache License, Version 2.0 (the "License");
   5  * you may not use this file except in compliance with the License.
   6  * You may obtain a copy of the License at:
   7  *
   8  *     http://www.apache.org/licenses/LICENSE-2.0
   9  *
  10  * Unless required by applicable law or agreed to in writing, software
  11  * distributed under the License is distributed on an "AS IS" BASIS,
  12  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  13  * See the License for the specific language governing permissions and
  14  * limitations under the License.
  15  */
  16
  17 #include "precomp.h"
  18 #include "NetProto.h"
  19 #include "Switch.h"
  20 #include "Vport.h"
  21 #include "Flow.h"
  22 #include "Vxlan.h"
  23 #include "IpHelper.h"
  24 #include "Checksum.h"
  25 #include "User.h"
  26 #include "PacketIO.h"
  27 #include "Flow.h"
  28 #include "PacketParser.h"
  29
  30 #pragma warning( push )
  31 #pragma warning( disable:4127 )
  32
  33
  34 #ifdef OVS_DBG_MOD
  35 #undef OVS_DBG_MOD
  36 #endif
  37 #define OVS_DBG_MOD OVS_DBG_VXLAN
  38 #include "Debug.h"
  39
  40 /* Helper macro to check if a VXLAN ID is valid. */
  41 #define VXLAN_ID_IS_VALID(vxlanID) (0 < (vxlanID) && (vxlanID) <= 0xffffff)
  42 #define VXLAN_TUNNELID_TO_VNI(_tID)   (UINT32)(((UINT64)(_tID)) >> 40)
  43 #define VXLAN_VNI_TO_TUNNELID(_vni) (((UINT64)(_vni)) << 40)
  44 #define IP_DF_NBO 0x0040
  45 #define VXLAN_DEFAULT_TTL 64
  46 #define VXLAN_MULTICAST_TTL 64
  47 #define VXLAN_DEFAULT_INSTANCE_ID 1
  48
  49 /* Move to a header file */
  50 extern POVS_SWITCH_CONTEXT gOvsSwitchContext;
  51
  52 /*
  53  *----------------------------------------------------------------------------
  54  * This function verifies if the VXLAN tunnel already exists, in order to
  55  * avoid sending a duplicate request to the WFP base filtering engine.
  56  *----------------------------------------------------------------------------
  57  */
  58 static BOOLEAN
  59 OvsIsTunnelFilterCreated(POVS_SWITCH_CONTEXT switchContext,
  60                          UINT16 udpPortDest)
  61 {
  62     for (UINT hash = 0; hash < OVS_MAX_VPORT_ARRAY_SIZE; hash++) {
  63         PLIST_ENTRY head, link, next;
  64
  65         head = &(switchContext->portNoHashArray[hash & OVS_VPORT_MASK]);
  66         LIST_FORALL_SAFE(head, link, next) {
  67             POVS_VPORT_ENTRY vport = NULL;
  68             POVS_VXLAN_VPORT vxlanPort = NULL;
  69             vport = CONTAINING_RECORD(link, OVS_VPORT_ENTRY, portNoLink);
  70             vxlanPort = (POVS_VXLAN_VPORT)vport->priv;
  71             if (vxlanPort) {
  72                 if ((udpPortDest == vxlanPort->dstPort)) {
  73                     /* The VXLAN tunnel was already created. */
  74                     return TRUE;
  75                 }
  76             }
  77         }
  78     }
  79
  80     return FALSE;
  81 }
  82
  83 /*
  84  *----------------------------------------------------------------------------
  85  * This function allocates and initializes the OVS_VXLAN_VPORT. The function
  86  * also creates a WFP tunnel filter for the necessary destination port. The
  87  * tunnel filter create request is passed to the tunnel filter threads that
  88  * will complete the request at a later time when IRQL is lowered to
  89  * PASSIVE_LEVEL.
  90  *
  91  * udpDestPort: the vxlan is set as payload to a udp frame. If the destination
  92  * port of an udp frame is udpDestPort, we understand it to be vxlan.
  93  *----------------------------------------------------------------------------
  94  */
  95 NTSTATUS
  96 OvsInitVxlanTunnel(PIRP irp,
  97                    POVS_VPORT_ENTRY vport,
  98                    UINT16 udpDestPort,
  99                    PFNTunnelVportPendingOp callback,
 100                    PVOID tunnelContext)
 101 {
 102     NTSTATUS status = STATUS_SUCCESS;
 103     POVS_VXLAN_VPORT vxlanPort = NULL;
 104
 105     vxlanPort = OvsAllocateMemoryWithTag(sizeof (*vxlanPort),
 106                                          OVS_VXLAN_POOL_TAG);
 107     if (vxlanPort == NULL) {
 108         return STATUS_INSUFFICIENT_RESOURCES;
 109     }
 110
 111     RtlZeroMemory(vxlanPort, sizeof(*vxlanPort));
 112     vxlanPort->dstPort = udpDestPort;
 113     vport->priv = (PVOID)vxlanPort;
 114
 115     if (!OvsIsTunnelFilterCreated(gOvsSwitchContext, udpDestPort)) {
 116         status = OvsTunelFilterCreate(irp,
 117                                       udpDestPort,
 118                                       &vxlanPort->filterID,
 119                                       callback,
 120                                       tunnelContext);
 121     } else {
 122         status = STATUS_OBJECT_NAME_EXISTS;
 123     }
 124
 125     return status;
 126 }
 127
 128 /*
 129  *----------------------------------------------------------------------------
 130  * This function releases the OVS_VXLAN_VPORT. The function also deletes the
 131  * WFP tunnel filter previously created. The tunnel filter delete request is
 132  * passed to the tunnel filter threads that will complete the request at a
 133  * later time when IRQL is lowered to PASSIVE_LEVEL.
 134  *----------------------------------------------------------------------------
 135  */
 136 NTSTATUS
 137 OvsCleanupVxlanTunnel(PIRP irp,
 138                       POVS_VPORT_ENTRY vport,
 139                       PFNTunnelVportPendingOp callback,
 140                       PVOID tunnelContext)
 141 {
 142     NTSTATUS status = STATUS_SUCCESS;
 143     POVS_VXLAN_VPORT vxlanPort = NULL;
 144
 145     if (vport->ovsType != OVS_VPORT_TYPE_VXLAN ||
 146         vport->priv == NULL) {
 147         return STATUS_SUCCESS;
 148     }
 149
 150     vxlanPort = (POVS_VXLAN_VPORT)vport->priv;
 151
 152     if (vxlanPort->filterID != 0) {
 153         status = OvsTunelFilterDelete(irp,
 154                                       vxlanPort->filterID,
 155                                       callback,
 156                                       tunnelContext);
 157     }
 158
 159     OvsFreeMemoryWithTag(vport->priv, OVS_VXLAN_POOL_TAG);
 160     vport->priv = NULL;
 161
 162     return status;
 163 }
 164
 165
 166 /*
 167  *----------------------------------------------------------------------------
 168  * OvsDoEncapVxlan
 169  *     Encapsulates the packet.
 170  *----------------------------------------------------------------------------
 171  */
 172 static __inline NDIS_STATUS
 173 OvsDoEncapVxlan(POVS_VPORT_ENTRY vport,
 174                 PNET_BUFFER_LIST curNbl,
 175                 OvsIPv4TunnelKey *tunKey,
 176                 POVS_FWD_INFO fwdInfo,
 177                 POVS_PACKET_HDR_INFO layers,
 178                 POVS_SWITCH_CONTEXT switchContext,
 179                 PNET_BUFFER_LIST *newNbl)
 180 {
 181     NDIS_STATUS status;
 182     PNET_BUFFER curNb;
 183     PMDL curMdl;
 184     PUINT8 bufferStart;
 185     EthHdr *ethHdr;
 186     IPHdr *ipHdr;
 187     UDPHdr *udpHdr;
 188     VXLANHdr *vxlanHdr;
 189     POVS_VXLAN_VPORT vportVxlan;
 190     UINT32 headRoom = OvsGetVxlanTunHdrSize();
 191     UINT32 packetLength;
 192
 193     /*
 194      * XXX: the assumption currently is that the NBL is owned by OVS, and
 195      * headroom has already been allocated as part of allocating the NBL and
 196      * MDL.
 197      */
 198     curNb = NET_BUFFER_LIST_FIRST_NB(curNbl);
 199     packetLength = NET_BUFFER_DATA_LENGTH(curNb);
 200     if (layers->isTcp) {
 201         NDIS_TCP_LARGE_SEND_OFFLOAD_NET_BUFFER_LIST_INFO tsoInfo;
 202
 203         tsoInfo.Value = NET_BUFFER_LIST_INFO(curNbl,
 204                 TcpLargeSendNetBufferListInfo);
 205         OVS_LOG_TRACE("MSS %u packet len %u", tsoInfo.LsoV1Transmit.MSS, packetLength);
 206         if (tsoInfo.LsoV1Transmit.MSS) {
 207             OVS_LOG_TRACE("l4Offset %d", layers->l4Offset);
 208             *newNbl = OvsTcpSegmentNBL(switchContext, curNbl, layers,
 209                         tsoInfo.LsoV1Transmit.MSS, headRoom);
 210             if (*newNbl == NULL) {
 211                 OVS_LOG_ERROR("Unable to segment NBL");
 212                 return NDIS_STATUS_FAILURE;
 213             }
 214         }
 215     }
 216
 217     vportVxlan = (POVS_VXLAN_VPORT) GetOvsVportPriv(vport);
 218     ASSERT(vportVxlan);
 219
 220     /* If we didn't split the packet above, make a copy now */
 221     if (*newNbl == NULL) {
 222         *newNbl = OvsPartialCopyNBL(switchContext, curNbl, 0, headRoom,
 223                                     FALSE /*NBL info*/);
 224         if (*newNbl == NULL) {
 225             OVS_LOG_ERROR("Unable to copy NBL");
 226             return NDIS_STATUS_FAILURE;
 227         }
 228     }
 229
 230     curNbl = *newNbl;
 231     for (curNb = NET_BUFFER_LIST_FIRST_NB(curNbl); curNb != NULL;
 232             curNb = curNb->Next) {
 233         status = NdisRetreatNetBufferDataStart(curNb, headRoom, 0, NULL);
 234         if (status != NDIS_STATUS_SUCCESS) {
 235             goto ret_error;
 236         }
 237
 238         curMdl = NET_BUFFER_CURRENT_MDL(curNb);
 239         bufferStart = (PUINT8)MmGetSystemAddressForMdlSafe(curMdl, LowPagePriority);
 240         if (!bufferStart) {
 241             status = NDIS_STATUS_RESOURCES;
 242             goto ret_error;
 243         }
 244
 245         bufferStart += NET_BUFFER_CURRENT_MDL_OFFSET(curNb);
 246         if (NET_BUFFER_NEXT_NB(curNb)) {
 247             OVS_LOG_TRACE("nb length %u next %u", NET_BUFFER_DATA_LENGTH(curNb),
 248                           NET_BUFFER_DATA_LENGTH(curNb->Next));
 249         }
 250
 251         /* L2 header */
 252         ethHdr = (EthHdr *)bufferStart;
 253         ASSERT(((PCHAR)&fwdInfo->dstMacAddr + sizeof fwdInfo->dstMacAddr) ==
 254                (PCHAR)&fwdInfo->srcMacAddr);
 255         NdisMoveMemory(ethHdr->Destination, fwdInfo->dstMacAddr,
 256                        sizeof ethHdr->Destination + sizeof ethHdr->Source);
 257         ethHdr->Type = htons(ETH_TYPE_IPV4);
 258
 259         // XXX: question: there are fields in the OvsIPv4TunnelKey for ttl and such,
 260         // should we use those values instead? or will they end up being
 261         // uninitialized;
 262         /* IP header */
 263         ipHdr = (IPHdr *)((PCHAR)ethHdr + sizeof *ethHdr);
 264
 265         ipHdr->ihl = sizeof *ipHdr / 4;
 266         ipHdr->version = IPV4;
 267         ipHdr->tos = 0;
 268         ipHdr->tot_len = htons(NET_BUFFER_DATA_LENGTH(curNb) - sizeof *ethHdr);
 269         ipHdr->id = 0;
 270         ipHdr->frag_off = IP_DF_NBO;
 271         ipHdr->ttl = tunKey->ttl ? tunKey->ttl : VXLAN_DEFAULT_TTL;
 272         ipHdr->protocol = IPPROTO_UDP;
 273         ASSERT(tunKey->dst == fwdInfo->dstIpAddr);
 274         ASSERT(tunKey->src == fwdInfo->srcIpAddr || tunKey->src == 0);
 275         ipHdr->saddr = fwdInfo->srcIpAddr;
 276         ipHdr->daddr = fwdInfo->dstIpAddr;
 277         ipHdr->check = 0;
 278         ipHdr->check = IPChecksum((UINT8 *)ipHdr, sizeof *ipHdr, 0);
 279
 280         /* UDP header */
 281         udpHdr = (UDPHdr *)((PCHAR)ipHdr + sizeof *ipHdr);
 282         udpHdr->source = htons(tunKey->flow_hash | 32768);
 283         udpHdr->dest = htons(vportVxlan->dstPort);
 284         udpHdr->len = htons(NET_BUFFER_DATA_LENGTH(curNb) - headRoom +
 285                             sizeof *udpHdr + sizeof *vxlanHdr);
 286         udpHdr->check = 0;
 287
 288         /* VXLAN header */
 289         vxlanHdr = (VXLANHdr *)((PCHAR)udpHdr + sizeof *udpHdr);
 290         vxlanHdr->flags1 = 0;
 291         vxlanHdr->locallyReplicate = 0;
 292         vxlanHdr->flags2 = 0;
 293         vxlanHdr->reserved1 = 0;
 294         if (tunKey->flags | OVS_TNL_F_KEY) {
 295             vxlanHdr->vxlanID = VXLAN_TUNNELID_TO_VNI(tunKey->tunnelId);
 296             vxlanHdr->instanceID = 1;
 297         }
 298         vxlanHdr->reserved2 = 0;
 299     }
 300     return STATUS_SUCCESS;
 301
 302 ret_error:
 303     OvsCompleteNBL(switchContext, *newNbl, TRUE);
 304     *newNbl = NULL;
 305     return status;
 306 }
 307
 308
 309 /*
 310  *----------------------------------------------------------------------------
 311  * OvsEncapVxlan --
 312  *     Encapsulates the packet if L2/L3 for destination resolves. Otherwise,
 313  *     enqueues a callback that does encapsulatation after resolution.
 314  *----------------------------------------------------------------------------
 315  */
 316 NDIS_STATUS
 317 OvsEncapVxlan(POVS_VPORT_ENTRY vport,
 318               PNET_BUFFER_LIST curNbl,
 319               OvsIPv4TunnelKey *tunKey,
 320               POVS_SWITCH_CONTEXT switchContext,
 321               POVS_PACKET_HDR_INFO layers,
 322               PNET_BUFFER_LIST *newNbl)
 323 {
 324     NTSTATUS status;
 325     OVS_FWD_INFO fwdInfo;
 326
 327     status = OvsLookupIPFwdInfo(tunKey->dst, &fwdInfo);
 328     if (status != STATUS_SUCCESS) {
 329         OvsFwdIPHelperRequest(NULL, 0, tunKey, NULL, NULL, NULL);
 330         // return NDIS_STATUS_PENDING;
 331         /*
 332          * XXX: Don't know if the completionList will make any sense when
 333          * accessed in the callback. Make sure the caveats are known.
 334          *
 335          * XXX: This code will work once we are able to grab locks in the
 336          * callback.
 337          */
 338         return NDIS_STATUS_FAILURE;
 339     }
 340
 341     return OvsDoEncapVxlan(vport, curNbl, tunKey, &fwdInfo, layers,
 342                            switchContext, newNbl);
 343 }
 344
 345 /*
 346  *----------------------------------------------------------------------------
 347  * OvsCalculateUDPChecksum
 348  *     Calculate UDP checksum
 349  *----------------------------------------------------------------------------
 350  */
 351 static __inline NDIS_STATUS
 352 OvsCalculateUDPChecksum(PNET_BUFFER_LIST curNbl,
 353                         PNET_BUFFER curNb,
 354                         IPHdr *ipHdr,
 355                         UDPHdr *udpHdr,
 356                         UINT32 packetLength)
 357 {
 358     NDIS_TCP_IP_CHECKSUM_NET_BUFFER_LIST_INFO csumInfo;
 359     UINT16 checkSum;
 360
 361     csumInfo.Value = NET_BUFFER_LIST_INFO(curNbl, TcpIpChecksumNetBufferListInfo);
 362
 363     /* Next check if UDP checksum has been calculated. */
 364     if (!csumInfo.Receive.UdpChecksumSucceeded) {
 365         UINT32 l4Payload;
 366
 367         checkSum = udpHdr->check;
 368
 369         l4Payload = packetLength - sizeof(EthHdr) - ipHdr->ihl * 4;
 370         udpHdr->check = 0;
 371         udpHdr->check =
 372             IPPseudoChecksum((UINT32 *)&ipHdr->saddr,
 373                              (UINT32 *)&ipHdr->daddr,
 374                              IPPROTO_UDP, (UINT16)l4Payload);
 375         udpHdr->check = CalculateChecksumNB(curNb, (UINT16)l4Payload,
 376             sizeof(EthHdr) + ipHdr->ihl * 4);
 377         if (checkSum != udpHdr->check) {
 378             OVS_LOG_TRACE("UDP checksum incorrect.");
 379             return NDIS_STATUS_INVALID_PACKET;
 380         }
 381     }
 382
 383     csumInfo.Receive.UdpChecksumSucceeded = 1;
 384     NET_BUFFER_LIST_INFO(curNbl, TcpIpChecksumNetBufferListInfo) = csumInfo.Value;
 385     return NDIS_STATUS_SUCCESS;
 386 }
 387
 388 /*
 389  *----------------------------------------------------------------------------
 390  * OvsDecapVxlan
 391  *     Decapsulates to tunnel header in 'curNbl' and puts into 'tunKey'.
 392  *----------------------------------------------------------------------------
 393  */
 394 NDIS_STATUS
 395 OvsDecapVxlan(POVS_SWITCH_CONTEXT switchContext,
 396               PNET_BUFFER_LIST curNbl,
 397               OvsIPv4TunnelKey *tunKey,
 398               PNET_BUFFER_LIST *newNbl)
 399 {
 400     PNET_BUFFER curNb;
 401     PMDL curMdl;
 402     EthHdr *ethHdr;
 403     IPHdr *ipHdr;
 404     UDPHdr *udpHdr;
 405     VXLANHdr *vxlanHdr;
 406     UINT32 tunnelSize = 0, packetLength = 0;
 407     PUINT8 bufferStart;
 408     NDIS_STATUS status;
 409
 410     /* Check the the length of the UDP payload */
 411     curNb = NET_BUFFER_LIST_FIRST_NB(curNbl);
 412     packetLength = NET_BUFFER_DATA_LENGTH(curNb);
 413     tunnelSize = OvsGetVxlanTunHdrSize();
 414     if (packetLength <= tunnelSize) {
 415         return NDIS_STATUS_INVALID_LENGTH;
 416     }
 417
 418     /*
 419      * Create a copy of the NBL so that we have all the headers in one MDL.
 420      */
 421     *newNbl = OvsPartialCopyNBL(switchContext, curNbl,
 422                                 tunnelSize + OVS_DEFAULT_COPY_SIZE, 0,
 423                                 TRUE /*copy NBL info */);
 424
 425     if (*newNbl == NULL) {
 426         return NDIS_STATUS_RESOURCES;
 427     }
 428
 429     /* XXX: Handle VLAN header. */
 430     curNbl = *newNbl;
 431     curNb = NET_BUFFER_LIST_FIRST_NB(curNbl);
 432     curMdl = NET_BUFFER_CURRENT_MDL(curNb);
 433     bufferStart = (PUINT8)MmGetSystemAddressForMdlSafe(curMdl, LowPagePriority) +
 434                   NET_BUFFER_CURRENT_MDL_OFFSET(curNb);
 435     if (!bufferStart) {
 436         status = NDIS_STATUS_RESOURCES;
 437         goto dropNbl;
 438     }
 439
 440     ethHdr = (EthHdr *)bufferStart;
 441     /* XXX: Handle IP options. */
 442     ipHdr = (IPHdr *)((PCHAR)ethHdr + sizeof *ethHdr);
 443     tunKey->src = ipHdr->saddr;
 444     tunKey->dst = ipHdr->daddr;
 445     tunKey->tos = ipHdr->tos;
 446     tunKey->ttl = ipHdr->ttl;
 447     tunKey->pad = 0;
 448     udpHdr = (UDPHdr *)((PCHAR)ipHdr + sizeof *ipHdr);
 449
 450     /* Validate if NIC has indicated checksum failure. */
 451     status = OvsValidateUDPChecksum(curNbl, udpHdr->check == 0);
 452     if (status != NDIS_STATUS_SUCCESS) {
 453         goto dropNbl;
 454     }
 455
 456     /* Calculate and verify UDP checksum if NIC didn't do it. */
 457     if (udpHdr->check != 0) {
 458         status = OvsCalculateUDPChecksum(curNbl, curNb, ipHdr, udpHdr, packetLength);
 459         if (status != NDIS_STATUS_SUCCESS) {
 460             goto dropNbl;
 461         }
 462     }
 463
 464     vxlanHdr = (VXLANHdr *)((PCHAR)udpHdr + sizeof *udpHdr);
 465     if (vxlanHdr->instanceID) {
 466         tunKey->flags = OVS_TNL_F_KEY;
 467         tunKey->tunnelId = VXLAN_VNI_TO_TUNNELID(vxlanHdr->vxlanID);
 468     } else {
 469         tunKey->flags = 0;
 470         tunKey->tunnelId = 0;
 471     }
 472
 473     /* Clear out the receive flag for the inner packet. */
 474     NET_BUFFER_LIST_INFO(curNbl, TcpIpChecksumNetBufferListInfo) = 0;
 475     NdisAdvanceNetBufferDataStart(curNb, tunnelSize, FALSE, NULL);
 476     return NDIS_STATUS_SUCCESS;
 477
 478 dropNbl:
 479     OvsCompleteNBL(switchContext, *newNbl, TRUE);
 480     *newNbl = NULL;
 481     return status;
 482 }
 483
 484
 485 NDIS_STATUS
 486 OvsSlowPathDecapVxlan(const PNET_BUFFER_LIST packet,
 487                    OvsIPv4TunnelKey *tunnelKey)
 488 {
 489     NDIS_STATUS status = NDIS_STATUS_FAILURE;
 490     UDPHdr udpStorage;
 491     const UDPHdr *udp;
 492     VXLANHdr *VxlanHeader;
 493     VXLANHdr  VxlanHeaderBuffer;
 494     struct IPHdr ip_storage;
 495     const struct IPHdr *nh;
 496     OVS_PACKET_HDR_INFO layers;
 497
 498     layers.value = 0;
 499
 500     do {
 501         nh = OvsGetIp(packet, layers.l3Offset, &ip_storage);
 502         if (nh) {
 503             layers.l4Offset = layers.l3Offset + nh->ihl * 4;
 504         } else {
 505             break;
 506         }
 507
 508         /* make sure it's a VXLAN packet */
 509         udp = OvsGetUdp(packet, layers.l4Offset, &udpStorage);
 510         if (udp) {
 511             layers.l7Offset = layers.l4Offset + sizeof *udp;
 512         } else {
 513             break;
 514         }
 515
 516         VxlanHeader = (VXLANHdr *)OvsGetPacketBytes(packet,
 517                                                     sizeof(*VxlanHeader),
 518                                                     layers.l7Offset,
 519                                                     &VxlanHeaderBuffer);
 520
 521         if (VxlanHeader) {
 522             tunnelKey->src = nh->saddr;
 523             tunnelKey->dst = nh->daddr;
 524             tunnelKey->ttl = nh->ttl;
 525             tunnelKey->tos = nh->tos;
 526             if (VxlanHeader->instanceID) {
 527                 tunnelKey->flags = OVS_TNL_F_KEY;
 528                 tunnelKey->tunnelId = VXLAN_VNI_TO_TUNNELID(VxlanHeader->vxlanID);
 529             } else {
 530                 tunnelKey->flags = 0;
 531                 tunnelKey->tunnelId = 0;
 532             }
 533         } else {
 534             break;
 535         }
 536         status = NDIS_STATUS_SUCCESS;
 537
 538     } while(FALSE);
 539
 540     return status;
 541 }
 542
 543 #pragma warning( pop )