datapath-windows/ovsext/OvsVxlan.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 "OvsNetProto.h"
  19 #include "OvsSwitch.h"
  20 #include "OvsVport.h"
  21 #include "OvsFlow.h"
  22 #include "OvsVxlan.h"
  23 #include "OvsIpHelper.h"
  24 #include "OvsChecksum.h"
  25 #include "OvsUser.h"
  26 #include "OvsPacketIO.h"
  27 #include "OvsFlow.h"
  28 #include "OvsPacketParser.h"
  29 #include "OvsChecksum.h"
  30
  31 #pragma warning( push )
  32 #pragma warning( disable:4127 )
  33
  34
  35 #ifdef OVS_DBG_MOD
  36 #undef OVS_DBG_MOD
  37 #endif
  38 #define OVS_DBG_MOD OVS_DBG_VXLAN
  39 #include "OvsDebug.h"
  40
  41 /* Helper macro to check if a VXLAN ID is valid. */
  42 #define VXLAN_ID_IS_VALID(vxlanID) (0 < (vxlanID) && (vxlanID) <= 0xffffff)
  43 #define VXLAN_TUNNELID_TO_VNI(_tID)   (UINT32)(((UINT64)(_tID)) >> 40)
  44 #define VXLAN_VNI_TO_TUNNELID(_vni) (((UINT64)(_vni)) << 40)
  45 #define IP_DF_NBO 0x0040
  46 #define VXLAN_DEFAULT_TTL 64
  47 #define VXLAN_MULTICAST_TTL 64
  48 #define VXLAN_DEFAULT_INSTANCE_ID 1
  49
  50 /* Move to a header file */
  51 extern POVS_SWITCH_CONTEXT gOvsSwitchContext;
  52
  53 NTSTATUS
  54 OvsInitVxlanTunnel(POVS_VPORT_ENTRY vport,
  55                    POVS_VPORT_ADD_REQUEST addReq)
  56 {
  57     POVS_VXLAN_VPORT vxlanPort;
  58     NTSTATUS status = STATUS_SUCCESS;
  59
  60     ASSERT(addReq->type == OVSWIN_VPORT_TYPE_VXLAN);
  61
  62     vxlanPort = OvsAllocateMemory(sizeof (*vxlanPort));
  63     if (vxlanPort == NULL) {
  64         status =  STATUS_INSUFFICIENT_RESOURCES;
  65     } else {
  66         RtlZeroMemory(vxlanPort, sizeof (*vxlanPort));
  67         vxlanPort->dstPort = addReq->dstPort;
  68         /*
  69          * since we are installing the WFP filter before the port is created
  70          * We need to check if it is the same number
  71          * XXX should be removed later
  72          */
  73         ASSERT(vxlanPort->dstPort == VXLAN_UDP_PORT);
  74         vport->priv = (PVOID)vxlanPort;
  75     }
  76     return status;
  77 }
  78
  79
  80 VOID
  81 OvsCleanupVxlanTunnel(POVS_VPORT_ENTRY vport)
  82 {
  83     if (vport->ovsType != OVSWIN_VPORT_TYPE_VXLAN ||
  84         vport->priv == NULL) {
  85         return;
  86     }
  87
  88     OvsFreeMemory(vport->priv);
  89     vport->priv = NULL;
  90 }
  91
  92
  93 /*
  94  *----------------------------------------------------------------------------
  95  * OvsDoEncapVxlan
  96  *     Encapsulates the packet.
  97  *----------------------------------------------------------------------------
  98  */
  99 static __inline NDIS_STATUS
 100 OvsDoEncapVxlan(PNET_BUFFER_LIST curNbl,
 101                 OvsIPv4TunnelKey *tunKey,
 102                 POVS_FWD_INFO fwdInfo,
 103                 POVS_PACKET_HDR_INFO layers,
 104                 POVS_SWITCH_CONTEXT switchContext,
 105                 PNET_BUFFER_LIST *newNbl)
 106 {
 107     NDIS_STATUS status;
 108     PNET_BUFFER curNb;
 109     PMDL curMdl;
 110     PUINT8 bufferStart;
 111     EthHdr *ethHdr;
 112     IPHdr *ipHdr;
 113     UDPHdr *udpHdr;
 114     VXLANHdr *vxlanHdr;
 115     UINT32 headRoom = OvsGetVxlanTunHdrSize();
 116     UINT32 packetLength;
 117
 118     /*
 119      * XXX: the assumption currently is that the NBL is owned by OVS, and
 120      * headroom has already been allocated as part of allocating the NBL and
 121      * MDL.
 122      */
 123     curNb = NET_BUFFER_LIST_FIRST_NB(curNbl);
 124     packetLength = NET_BUFFER_DATA_LENGTH(curNb);
 125     if (layers->isTcp) {
 126         NDIS_TCP_LARGE_SEND_OFFLOAD_NET_BUFFER_LIST_INFO tsoInfo;
 127
 128         tsoInfo.Value = NET_BUFFER_LIST_INFO(curNbl,
 129                 TcpLargeSendNetBufferListInfo);
 130         OVS_LOG_TRACE("MSS %u packet len %u", tsoInfo.LsoV1Transmit.MSS, packetLength);
 131         if (tsoInfo.LsoV1Transmit.MSS) {
 132             OVS_LOG_TRACE("l4Offset %d", layers->l4Offset);
 133             *newNbl = OvsTcpSegmentNBL(switchContext, curNbl, layers,
 134                         tsoInfo.LsoV1Transmit.MSS, headRoom);
 135             if (*newNbl == NULL) {
 136                 OVS_LOG_ERROR("Unable to segment NBL");
 137                 return NDIS_STATUS_FAILURE;
 138             }
 139         }
 140     }
 141     /* If we didn't split the packet above, make a copy now */
 142     if (*newNbl == NULL) {
 143         *newNbl = OvsPartialCopyNBL(switchContext, curNbl, 0, headRoom,
 144                                     FALSE /*NBL info*/);
 145         if (*newNbl == NULL) {
 146             OVS_LOG_ERROR("Unable to copy NBL");
 147             return NDIS_STATUS_FAILURE;
 148         }
 149     }
 150
 151     curNbl = *newNbl;
 152     for (curNb = NET_BUFFER_LIST_FIRST_NB(curNbl); curNb != NULL;
 153             curNb = curNb->Next) {
 154         status = NdisRetreatNetBufferDataStart(curNb, headRoom, 0, NULL);
 155         if (status != NDIS_STATUS_SUCCESS) {
 156             goto ret_error;
 157         }
 158
 159         curMdl = NET_BUFFER_CURRENT_MDL(curNb);
 160         bufferStart = (PUINT8)MmGetSystemAddressForMdlSafe(curMdl, LowPagePriority);
 161         if (!bufferStart) {
 162             status = NDIS_STATUS_RESOURCES;
 163             goto ret_error;
 164         }
 165
 166         bufferStart += NET_BUFFER_CURRENT_MDL_OFFSET(curNb);
 167         if (NET_BUFFER_NEXT_NB(curNb)) {
 168             OVS_LOG_TRACE("nb length %u next %u", NET_BUFFER_DATA_LENGTH(curNb),
 169                           NET_BUFFER_DATA_LENGTH(curNb->Next));
 170         }
 171
 172         /* L2 header */
 173         ethHdr = (EthHdr *)bufferStart;
 174         NdisMoveMemory(ethHdr->Destination, fwdInfo->dstMacAddr,
 175                        sizeof ethHdr->Destination + sizeof ethHdr->Source);
 176         ASSERT(((PCHAR)&fwdInfo->dstMacAddr + sizeof fwdInfo->dstMacAddr) ==
 177                (PCHAR)&fwdInfo->srcMacAddr);
 178         ethHdr->Type = htons(ETH_TYPE_IPV4);
 179
 180         // XXX: question: there are fields in the OvsIPv4TunnelKey for ttl and such,
 181         // should we use those values instead? or will they end up being
 182         // uninitialized;
 183         /* IP header */
 184         ipHdr = (IPHdr *)((PCHAR)ethHdr + sizeof *ethHdr);
 185
 186         ipHdr->ihl = sizeof *ipHdr / 4;
 187         ipHdr->version = IPV4;
 188         ipHdr->tos = 0;
 189         ipHdr->tot_len = htons(NET_BUFFER_DATA_LENGTH(curNb) - sizeof *ethHdr);
 190         ipHdr->id = 0;
 191         ipHdr->frag_off = IP_DF_NBO;
 192         ipHdr->ttl = tunKey->ttl ? tunKey->ttl : VXLAN_DEFAULT_TTL;
 193         ipHdr->protocol = IPPROTO_UDP;
 194         ASSERT(tunKey->dst == fwdInfo->dstIpAddr);
 195         ASSERT(tunKey->src == fwdInfo->srcIpAddr || tunKey->src == 0);
 196         ipHdr->saddr = fwdInfo->srcIpAddr;
 197         ipHdr->daddr = fwdInfo->dstIpAddr;
 198         ipHdr->check = 0;
 199         ipHdr->check = IPChecksum((UINT8 *)ipHdr, sizeof *ipHdr, 0);
 200
 201         /* UDP header */
 202         udpHdr = (UDPHdr *)((PCHAR)ipHdr + sizeof *ipHdr);
 203         udpHdr->source = htons(tunKey->flow_hash | 32768);
 204         udpHdr->dest = VXLAN_UDP_PORT_NBO;
 205         udpHdr->len = htons(NET_BUFFER_DATA_LENGTH(curNb) - headRoom +
 206                             sizeof *udpHdr + sizeof *vxlanHdr);
 207         udpHdr->check = 0;
 208
 209         /* VXLAN header */
 210         vxlanHdr = (VXLANHdr *)((PCHAR)udpHdr + sizeof *udpHdr);
 211         vxlanHdr->flags1 = 0;
 212         vxlanHdr->locallyReplicate = 0;
 213         vxlanHdr->flags2 = 0;
 214         vxlanHdr->reserved1 = 0;
 215         if (tunKey->flags | OVS_TNL_F_KEY) {
 216             vxlanHdr->vxlanID = VXLAN_TUNNELID_TO_VNI(tunKey->tunnelId);
 217             vxlanHdr->instanceID = 1;
 218         }
 219         vxlanHdr->reserved2 = 0;
 220     }
 221     return STATUS_SUCCESS;
 222
 223 ret_error:
 224     OvsCompleteNBL(switchContext, *newNbl, TRUE);
 225     *newNbl = NULL;
 226     return status;
 227 }
 228
 229
 230 /*
 231  *----------------------------------------------------------------------------
 232  * OvsEncapVxlan --
 233  *     Encapsulates the packet if L2/L3 for destination resolves. Otherwise,
 234  *     enqueues a callback that does encapsulatation after resolution.
 235  *----------------------------------------------------------------------------
 236  */
 237 NDIS_STATUS
 238 OvsEncapVxlan(PNET_BUFFER_LIST curNbl,
 239               OvsIPv4TunnelKey *tunKey,
 240               POVS_SWITCH_CONTEXT switchContext,
 241               VOID *completionList,
 242               POVS_PACKET_HDR_INFO layers,
 243               PNET_BUFFER_LIST *newNbl)
 244 {
 245     NTSTATUS status;
 246     OVS_FWD_INFO fwdInfo;
 247     UNREFERENCED_PARAMETER(completionList);
 248
 249     status = OvsLookupIPFwdInfo(tunKey->dst, &fwdInfo);
 250     if (status != STATUS_SUCCESS) {
 251         OvsFwdIPHelperRequest(NULL, 0, tunKey, NULL, NULL, NULL);
 252         // return NDIS_STATUS_PENDING;
 253         /*
 254          * XXX: Don't know if the completionList will make any sense when
 255          * accessed in the callback. Make sure the caveats are known.
 256          *
 257          * XXX: This code will work once we are able to grab locks in the
 258          * callback.
 259          */
 260         return NDIS_STATUS_FAILURE;
 261     }
 262
 263     return OvsDoEncapVxlan(curNbl, tunKey, &fwdInfo, layers,
 264                            switchContext, newNbl);
 265 }
 266
 267
 268 /*
 269  *----------------------------------------------------------------------------
 270  * OvsIpHlprCbVxlan --
 271  *     Callback function for IP helper.
 272  *     XXX: not used currently
 273  *----------------------------------------------------------------------------
 274  */
 275 static VOID
 276 OvsIpHlprCbVxlan(PNET_BUFFER_LIST curNbl,
 277                  UINT32 inPort,
 278                  OvsIPv4TunnelKey *tunKey,
 279                  PVOID cbData1,
 280                  PVOID cbData2,
 281                  NTSTATUS result,
 282                  POVS_FWD_INFO fwdInfo)
 283 {
 284     OVS_PACKET_HDR_INFO layers;
 285     OvsFlowKey key;
 286     NDIS_STATUS status;
 287     UNREFERENCED_PARAMETER(inPort);
 288
 289     status = OvsExtractFlow(curNbl, inPort, &key, &layers, NULL);
 290     if (result == STATUS_SUCCESS) {
 291         status = OvsDoEncapVxlan(curNbl, tunKey, fwdInfo, &layers,
 292                 (POVS_SWITCH_CONTEXT)cbData1, NULL);
 293     } else {
 294         status = NDIS_STATUS_FAILURE;
 295     }
 296
 297     if (status != NDIS_STATUS_SUCCESS) {
 298         // XXX: Free up the NBL;
 299         return;
 300     }
 301
 302     OvsLookupFlowOutput((POVS_SWITCH_CONTEXT)cbData1, cbData2, curNbl);
 303 }
 304
 305 /*
 306  *----------------------------------------------------------------------------
 307  * OvsCalculateUDPChecksum
 308  *     Calculate UDP checksum
 309  *----------------------------------------------------------------------------
 310  */
 311 static __inline NDIS_STATUS
 312 OvsCalculateUDPChecksum(PNET_BUFFER_LIST curNbl,
 313                         PNET_BUFFER curNb,
 314                         IPHdr *ipHdr,
 315                         UDPHdr *udpHdr,
 316                         UINT32 packetLength)
 317 {
 318     NDIS_TCP_IP_CHECKSUM_NET_BUFFER_LIST_INFO csumInfo;
 319     UINT16 checkSum;
 320
 321     csumInfo.Value = NET_BUFFER_LIST_INFO(curNbl, TcpIpChecksumNetBufferListInfo);
 322
 323     /* Next check if UDP checksum has been calculated. */
 324     if (!csumInfo.Receive.UdpChecksumSucceeded) {
 325         UINT32 l4Payload;
 326
 327         checkSum = udpHdr->check;
 328
 329         l4Payload = packetLength - sizeof(EthHdr) - ipHdr->ihl * 4;
 330         udpHdr->check = 0;
 331         udpHdr->check =
 332             IPPseudoChecksum((UINT32 *)&ipHdr->saddr,
 333                              (UINT32 *)&ipHdr->daddr,
 334                              IPPROTO_UDP, (UINT16)l4Payload);
 335         udpHdr->check = CalculateChecksumNB(curNb, (UINT16)l4Payload,
 336             sizeof(EthHdr) + ipHdr->ihl * 4);
 337         if (checkSum != udpHdr->check) {
 338             OVS_LOG_TRACE("UDP checksum incorrect.");
 339             return NDIS_STATUS_INVALID_PACKET;
 340         }
 341     }
 342
 343     csumInfo.Receive.UdpChecksumSucceeded = 1;
 344     NET_BUFFER_LIST_INFO(curNbl, TcpIpChecksumNetBufferListInfo) = csumInfo.Value;
 345     return NDIS_STATUS_SUCCESS;
 346 }
 347
 348 /*
 349  *----------------------------------------------------------------------------
 350  * OvsDoDecapVxlan
 351  *     Decapsulates to tunnel header in 'curNbl' and puts into 'tunKey'.
 352  *----------------------------------------------------------------------------
 353  */
 354 NDIS_STATUS
 355 OvsDoDecapVxlan(POVS_SWITCH_CONTEXT switchContext,
 356                 PNET_BUFFER_LIST curNbl,
 357                 OvsIPv4TunnelKey *tunKey,
 358                 PNET_BUFFER_LIST *newNbl)
 359 {
 360     PNET_BUFFER curNb;
 361     PMDL curMdl;
 362     EthHdr *ethHdr;
 363     IPHdr *ipHdr;
 364     UDPHdr *udpHdr;
 365     VXLANHdr *vxlanHdr;
 366     UINT32 tunnelSize = 0, packetLength = 0;
 367     PUINT8 bufferStart;
 368     NDIS_STATUS status;
 369
 370     /* Check the the length of the UDP payload */
 371     curNb = NET_BUFFER_LIST_FIRST_NB(curNbl);
 372     packetLength = NET_BUFFER_DATA_LENGTH(curNb);
 373     tunnelSize = OvsGetVxlanTunHdrSize();
 374     if (packetLength <= tunnelSize) {
 375         return NDIS_STATUS_INVALID_LENGTH;
 376     }
 377
 378     /*
 379      * Create a copy of the NBL so that we have all the headers in one MDL.
 380      */
 381     *newNbl = OvsPartialCopyNBL(switchContext, curNbl,
 382                                 tunnelSize + OVS_DEFAULT_COPY_SIZE, 0,
 383                                 TRUE /*copy NBL info */);
 384
 385     if (*newNbl == NULL) {
 386         return NDIS_STATUS_RESOURCES;
 387     }
 388
 389     /* XXX: Handle VLAN header. */
 390     curNbl = *newNbl;
 391     curNb = NET_BUFFER_LIST_FIRST_NB(curNbl);
 392     curMdl = NET_BUFFER_CURRENT_MDL(curNb);
 393     bufferStart = (PUINT8)MmGetSystemAddressForMdlSafe(curMdl, LowPagePriority) +
 394                   NET_BUFFER_CURRENT_MDL_OFFSET(curNb);
 395     if (!bufferStart) {
 396         status = NDIS_STATUS_RESOURCES;
 397         goto dropNbl;
 398     }
 399
 400     ethHdr = (EthHdr *)bufferStart;
 401     /* XXX: Handle IP options. */
 402     ipHdr = (IPHdr *)((PCHAR)ethHdr + sizeof *ethHdr);
 403     tunKey->src = ipHdr->saddr;
 404     tunKey->dst = ipHdr->daddr;
 405     tunKey->tos = ipHdr->tos;
 406     tunKey->ttl = ipHdr->ttl;
 407     tunKey->pad = 0;
 408     udpHdr = (UDPHdr *)((PCHAR)ipHdr + sizeof *ipHdr);
 409
 410     /* Validate if NIC has indicated checksum failure. */
 411     status = OvsValidateUDPChecksum(curNbl, udpHdr->check == 0);
 412     if (status != NDIS_STATUS_SUCCESS) {
 413         goto dropNbl;
 414     }
 415
 416     /* Calculate and verify UDP checksum if NIC didn't do it. */
 417     if (udpHdr->check != 0) {
 418         status = OvsCalculateUDPChecksum(curNbl, curNb, ipHdr, udpHdr, packetLength);
 419         if (status != NDIS_STATUS_SUCCESS) {
 420             goto dropNbl;
 421         }
 422     }
 423
 424     vxlanHdr = (VXLANHdr *)((PCHAR)udpHdr + sizeof *udpHdr);
 425     if (vxlanHdr->instanceID) {
 426         tunKey->flags = OVS_TNL_F_KEY;
 427         tunKey->tunnelId = VXLAN_VNI_TO_TUNNELID(vxlanHdr->vxlanID);
 428     } else {
 429         tunKey->flags = 0;
 430         tunKey->tunnelId = 0;
 431     }
 432
 433     /* Clear out the receive flag for the inner packet. */
 434     NET_BUFFER_LIST_INFO(curNbl, TcpIpChecksumNetBufferListInfo) = 0;
 435     NdisAdvanceNetBufferDataStart(curNb, tunnelSize, FALSE, NULL);
 436     return NDIS_STATUS_SUCCESS;
 437
 438 dropNbl:
 439     OvsCompleteNBL(switchContext, *newNbl, TRUE);
 440     *newNbl = NULL;
 441     return status;
 442 }
 443
 444
 445 NDIS_STATUS
 446 OvsSlowPathDecapVxlan(const PNET_BUFFER_LIST packet,
 447                    OvsIPv4TunnelKey *tunnelKey)
 448 {
 449     NDIS_STATUS status = NDIS_STATUS_FAILURE;
 450     UDPHdr udpStorage;
 451     const UDPHdr *udp;
 452     VXLANHdr *VxlanHeader;
 453     VXLANHdr  VxlanHeaderBuffer;
 454     struct IPHdr ip_storage;
 455     const struct IPHdr *nh;
 456     OVS_PACKET_HDR_INFO layers;
 457
 458     layers.value = 0;
 459
 460     do {
 461         nh = OvsGetIp(packet, layers.l3Offset, &ip_storage);
 462         if (nh) {
 463             layers.l4Offset = layers.l3Offset + nh->ihl * 4;
 464         } else {
 465             break;
 466         }
 467
 468         /* make sure it's a VXLAN packet */
 469         udp = OvsGetUdp(packet, layers.l4Offset, &udpStorage);
 470         if (udp) {
 471             layers.l7Offset = layers.l4Offset + sizeof *udp;
 472         } else {
 473             break;
 474         }
 475
 476         /* XXX Should be tested against the dynamic port # in the VXLAN vport */
 477         ASSERT(udp->dest == RtlUshortByteSwap(VXLAN_UDP_PORT));
 478
 479         VxlanHeader = (VXLANHdr *)OvsGetPacketBytes(packet,
 480                                                     sizeof(*VxlanHeader),
 481                                                     layers.l7Offset,
 482                                                     &VxlanHeaderBuffer);
 483
 484         if (VxlanHeader) {
 485             tunnelKey->src = nh->saddr;
 486             tunnelKey->dst = nh->daddr;
 487             tunnelKey->ttl = nh->ttl;
 488             tunnelKey->tos = nh->tos;
 489             if (VxlanHeader->instanceID) {
 490                 tunnelKey->flags = OVS_TNL_F_KEY;
 491                 tunnelKey->tunnelId = VXLAN_VNI_TO_TUNNELID(VxlanHeader->vxlanID);
 492             } else {
 493                 tunnelKey->flags = 0;
 494                 tunnelKey->tunnelId = 0;
 495             }
 496         } else {
 497             break;
 498         }
 499         status = NDIS_STATUS_SUCCESS;
 500
 501     } while(FALSE);
 502
 503     return status;
 504 }
 505
 506 #pragma warning( pop )