datapath-windows/ovsext/Vxlan.c

   1 /*
   2  * Copyright (c) 2014, 2016 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
  19 #include "Atomic.h"
  20 #include "Debug.h"
  21 #include "Flow.h"
  22 #include "IpHelper.h"
  23 #include "NetProto.h"
  24 #include "Offload.h"
  25 #include "PacketIO.h"
  26 #include "PacketParser.h"
  27 #include "Switch.h"
  28 #include "User.h"
  29 #include "Vport.h"
  30 #include "Vxlan.h"
  31
  32 #pragma warning( push )
  33 #pragma warning( disable:4127 )
  34
  35
  36 #ifdef OVS_DBG_MOD
  37 #undef OVS_DBG_MOD
  38 #endif
  39 #define OVS_DBG_MOD OVS_DBG_VXLAN
  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 /*
  54  *----------------------------------------------------------------------------
  55  * This function verifies if the VXLAN tunnel already exists, in order to
  56  * avoid sending a duplicate request to the WFP base filtering engine.
  57  *----------------------------------------------------------------------------
  58  */
  59 static BOOLEAN
  60 OvsIsTunnelFilterCreated(POVS_SWITCH_CONTEXT switchContext,
  61                          UINT16 udpPortDest)
  62 {
  63     for (UINT hash = 0; hash < OVS_MAX_VPORT_ARRAY_SIZE; hash++) {
  64         PLIST_ENTRY head, link, next;
  65
  66         head = &(switchContext->portNoHashArray[hash & OVS_VPORT_MASK]);
  67         LIST_FORALL_SAFE(head, link, next) {
  68             POVS_VPORT_ENTRY vport = NULL;
  69             POVS_VXLAN_VPORT vxlanPort = NULL;
  70             vport = CONTAINING_RECORD(link, OVS_VPORT_ENTRY, portNoLink);
  71             vxlanPort = (POVS_VXLAN_VPORT)vport->priv;
  72             if (vxlanPort) {
  73                 if ((udpPortDest == vxlanPort->dstPort)) {
  74                     /* The VXLAN tunnel was already created. */
  75                     return TRUE;
  76                 }
  77             }
  78         }
  79     }
  80
  81     return FALSE;
  82 }
  83
  84 /*
  85  *----------------------------------------------------------------------------
  86  * This function allocates and initializes the OVS_VXLAN_VPORT. The function
  87  * also creates a WFP tunnel filter for the necessary destination port. The
  88  * tunnel filter create request is passed to the tunnel filter threads that
  89  * will complete the request at a later time when IRQL is lowered to
  90  * PASSIVE_LEVEL.
  91  *
  92  * udpDestPort: the vxlan is set as payload to a udp frame. If the destination
  93  * port of an udp frame is udpDestPort, we understand it to be vxlan.
  94  *----------------------------------------------------------------------------
  95  */
  96 NTSTATUS
  97 OvsInitVxlanTunnel(PIRP irp,
  98                    POVS_VPORT_ENTRY vport,
  99                    UINT16 udpDestPort,
 100                    PFNTunnelVportPendingOp callback,
 101                    PVOID tunnelContext)
 102 {
 103     NTSTATUS status = STATUS_SUCCESS;
 104     POVS_VXLAN_VPORT vxlanPort = NULL;
 105
 106     vxlanPort = OvsAllocateMemoryWithTag(sizeof (*vxlanPort),
 107                                          OVS_VXLAN_POOL_TAG);
 108     if (vxlanPort == NULL) {
 109         return STATUS_INSUFFICIENT_RESOURCES;
 110     }
 111
 112     RtlZeroMemory(vxlanPort, sizeof(*vxlanPort));
 113     vxlanPort->dstPort = udpDestPort;
 114     vport->priv = (PVOID)vxlanPort;
 115
 116     if (!OvsIsTunnelFilterCreated(gOvsSwitchContext, udpDestPort)) {
 117         status = OvsTunnelFilterCreate(irp,
 118                                        udpDestPort,
 119                                        &vxlanPort->filterID,
 120                                        callback,
 121                                        tunnelContext);
 122     } else {
 123         status = STATUS_OBJECT_NAME_EXISTS;
 124     }
 125
 126     return status;
 127 }
 128
 129 /*
 130  *----------------------------------------------------------------------------
 131  * This function releases the OVS_VXLAN_VPORT. The function also deletes the
 132  * WFP tunnel filter previously created. The tunnel filter delete request is
 133  * passed to the tunnel filter threads that will complete the request at a
 134  * later time when IRQL is lowered to PASSIVE_LEVEL.
 135  *----------------------------------------------------------------------------
 136  */
 137 NTSTATUS
 138 OvsCleanupVxlanTunnel(PIRP irp,
 139                       POVS_VPORT_ENTRY vport,
 140                       PFNTunnelVportPendingOp callback,
 141                       PVOID tunnelContext)
 142 {
 143     NTSTATUS status = STATUS_SUCCESS;
 144     POVS_VXLAN_VPORT vxlanPort = NULL;
 145
 146     if (vport->ovsType != OVS_VPORT_TYPE_VXLAN ||
 147         vport->priv == NULL) {
 148         return STATUS_SUCCESS;
 149     }
 150
 151     vxlanPort = (POVS_VXLAN_VPORT)vport->priv;
 152
 153     if (vxlanPort->filterID != 0) {
 154         status = OvsTunnelFilterDelete(irp,
 155                                        vxlanPort->filterID,
 156                                        callback,
 157                                        tunnelContext);
 158     } else {
 159         OvsFreeMemoryWithTag(vport->priv, OVS_VXLAN_POOL_TAG);
 160         vport->priv = NULL;
 161     }
 162
 163     return status;
 164 }
 165
 166
 167 /*
 168  *----------------------------------------------------------------------------
 169  * OvsDoEncapVxlan
 170  *     Encapsulates the packet.
 171  *----------------------------------------------------------------------------
 172  */
 173 static __inline NDIS_STATUS
 174 OvsDoEncapVxlan(POVS_VPORT_ENTRY vport,
 175                 PNET_BUFFER_LIST curNbl,
 176                 OvsIPv4TunnelKey *tunKey,
 177                 POVS_FWD_INFO fwdInfo,
 178                 POVS_PACKET_HDR_INFO layers,
 179                 POVS_SWITCH_CONTEXT switchContext,
 180                 PNET_BUFFER_LIST *newNbl)
 181 {
 182     NDIS_STATUS status;
 183     PNET_BUFFER curNb;
 184     PMDL curMdl;
 185     PUINT8 bufferStart;
 186     EthHdr *ethHdr;
 187     IPHdr *ipHdr;
 188     UDPHdr *udpHdr;
 189     VXLANHdr *vxlanHdr;
 190     POVS_VXLAN_VPORT vportVxlan;
 191     UINT32 headRoom = OvsGetVxlanTunHdrSize();
 192     UINT32 packetLength;
 193     ULONG mss = 0;
 194     NDIS_TCP_IP_CHECKSUM_NET_BUFFER_LIST_INFO csumInfo;
 195
 196     curNb = NET_BUFFER_LIST_FIRST_NB(curNbl);
 197     packetLength = NET_BUFFER_DATA_LENGTH(curNb);
 198
 199     if (layers->isTcp) {
 200         mss = OVSGetTcpMSS(curNbl);
 201
 202         OVS_LOG_TRACE("MSS %u packet len %u", mss,
 203                       packetLength);
 204         if (mss) {
 205             OVS_LOG_TRACE("l4Offset %d", layers->l4Offset);
 206             *newNbl = OvsTcpSegmentNBL(switchContext, curNbl, layers,
 207                                        mss, headRoom, FALSE);
 208             if (*newNbl == NULL) {
 209                 OVS_LOG_ERROR("Unable to segment NBL");
 210                 return NDIS_STATUS_FAILURE;
 211             }
 212             /* Clear out LSO flags after this point */
 213             NET_BUFFER_LIST_INFO(*newNbl, TcpLargeSendNetBufferListInfo) = 0;
 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         csumInfo.Value = NET_BUFFER_LIST_INFO(curNbl,
 229                                               TcpIpChecksumNetBufferListInfo);
 230         status = OvsApplySWChecksumOnNB(layers, *newNbl, &csumInfo);
 231
 232         if (status != NDIS_STATUS_SUCCESS) {
 233             goto ret_error;
 234         }
 235     }
 236
 237     curNbl = *newNbl;
 238     for (curNb = NET_BUFFER_LIST_FIRST_NB(curNbl); curNb != NULL;
 239             curNb = curNb->Next) {
 240         status = NdisRetreatNetBufferDataStart(curNb, headRoom, 0, NULL);
 241         if (status != NDIS_STATUS_SUCCESS) {
 242             goto ret_error;
 243         }
 244
 245         curMdl = NET_BUFFER_CURRENT_MDL(curNb);
 246         bufferStart = (PUINT8)OvsGetMdlWithLowPriority(curMdl);
 247         if (!bufferStart) {
 248             status = NDIS_STATUS_RESOURCES;
 249             goto ret_error;
 250         }
 251
 252         bufferStart += NET_BUFFER_CURRENT_MDL_OFFSET(curNb);
 253         if (NET_BUFFER_NEXT_NB(curNb)) {
 254             OVS_LOG_TRACE("nb length %u next %u",
 255                           NET_BUFFER_DATA_LENGTH(curNb),
 256                           NET_BUFFER_DATA_LENGTH(curNb->Next));
 257         }
 258
 259         /* L2 header */
 260         ethHdr = (EthHdr *)bufferStart;
 261         NdisMoveMemory(ethHdr->Destination, fwdInfo->dstMacAddr,
 262                        sizeof ethHdr->Destination);
 263         NdisMoveMemory(ethHdr->Source, fwdInfo->srcMacAddr,
 264                        sizeof ethHdr->Source);
 265         ethHdr->Type = htons(ETH_TYPE_IPV4);
 266
 267         /* IP header */
 268         ipHdr = (IPHdr *)((PCHAR)ethHdr + sizeof *ethHdr);
 269
 270         ipHdr->ihl = sizeof *ipHdr / 4;
 271         ipHdr->version = IPPROTO_IPV4;
 272         ipHdr->tos = tunKey->tos;
 273         ipHdr->tot_len = htons(NET_BUFFER_DATA_LENGTH(curNb) - sizeof *ethHdr);
 274         ipHdr->id = (uint16)atomic_add64(&vportVxlan->ipId,
 275                                          NET_BUFFER_DATA_LENGTH(curNb));
 276         ipHdr->frag_off = (tunKey->flags & OVS_TNL_F_DONT_FRAGMENT) ?
 277                           IP_DF_NBO : 0;
 278         ipHdr->ttl = tunKey->ttl ? tunKey->ttl : VXLAN_DEFAULT_TTL;
 279         ipHdr->protocol = IPPROTO_UDP;
 280         ASSERT(tunKey->dst == fwdInfo->dstIpAddr);
 281         ASSERT(tunKey->src == fwdInfo->srcIpAddr || tunKey->src == 0);
 282         ipHdr->saddr = fwdInfo->srcIpAddr;
 283         ipHdr->daddr = fwdInfo->dstIpAddr;
 284
 285         ipHdr->check = 0;
 286
 287         /* UDP header */
 288         udpHdr = (UDPHdr *)((PCHAR)ipHdr + sizeof *ipHdr);
 289         udpHdr->source = htons(tunKey->flow_hash | MAXINT16);
 290         udpHdr->dest = tunKey->dst_port ? tunKey->dst_port :
 291                                           htons(vportVxlan->dstPort);
 292         udpHdr->len = htons(NET_BUFFER_DATA_LENGTH(curNb) - headRoom +
 293                             sizeof *udpHdr + sizeof *vxlanHdr);
 294
 295         if (tunKey->flags & OVS_TNL_F_CSUM) {
 296             udpHdr->check = IPPseudoChecksum(&ipHdr->saddr, &ipHdr->daddr,
 297                                              IPPROTO_UDP, ntohs(udpHdr->len));
 298         } else {
 299             udpHdr->check = 0;
 300         }
 301
 302         /* VXLAN header */
 303         vxlanHdr = (VXLANHdr *)((PCHAR)udpHdr + sizeof *udpHdr);
 304         vxlanHdr->flags1 = 0;
 305         vxlanHdr->locallyReplicate = 0;
 306         vxlanHdr->flags2 = 0;
 307         vxlanHdr->reserved1 = 0;
 308         vxlanHdr->vxlanID = VXLAN_TUNNELID_TO_VNI(tunKey->tunnelId);
 309         vxlanHdr->instanceID = 1;
 310         vxlanHdr->reserved2 = 0;
 311     }
 312
 313     csumInfo.Value = 0;
 314     csumInfo.Transmit.IpHeaderChecksum = 1;
 315     csumInfo.Transmit.IsIPv4 = 1;
 316     if (tunKey->flags & OVS_TNL_F_CSUM) {
 317         csumInfo.Transmit.UdpChecksum = 1;
 318     }
 319     NET_BUFFER_LIST_INFO(curNbl,
 320                          TcpIpChecksumNetBufferListInfo) = csumInfo.Value;
 321
 322
 323     return STATUS_SUCCESS;
 324
 325 ret_error:
 326     OvsCompleteNBL(switchContext, *newNbl, TRUE);
 327     *newNbl = NULL;
 328     return status;
 329 }
 330
 331
 332 /*
 333  *----------------------------------------------------------------------------
 334  * OvsEncapVxlan --
 335  *     Encapsulates the packet if L2/L3 for destination resolves. Otherwise,
 336  *     enqueues a callback that does encapsulation after resolution.
 337  *----------------------------------------------------------------------------
 338  */
 339 NDIS_STATUS
 340 OvsEncapVxlan(POVS_VPORT_ENTRY vport,
 341               PNET_BUFFER_LIST curNbl,
 342               OvsIPv4TunnelKey *tunKey,
 343               POVS_SWITCH_CONTEXT switchContext,
 344               POVS_PACKET_HDR_INFO layers,
 345               PNET_BUFFER_LIST *newNbl,
 346               POVS_FWD_INFO switchFwdInfo)
 347 {
 348     NTSTATUS status;
 349     OVS_FWD_INFO fwdInfo;
 350
 351     status = OvsLookupIPFwdInfo(tunKey->src, tunKey->dst, &fwdInfo);
 352     if (status != STATUS_SUCCESS) {
 353         OvsFwdIPHelperRequest(NULL, 0, tunKey, NULL, NULL, NULL);
 354         /*
 355          * XXX: Don't know if the completionList will make any sense when
 356          * accessed in the callback. Make sure the caveats are known.
 357          *
 358          * XXX: This code will work once we are able to grab locks in the
 359          * callback.
 360          */
 361         return NDIS_STATUS_FAILURE;
 362     }
 363
 364     RtlCopyMemory(switchFwdInfo->value, fwdInfo.value, sizeof fwdInfo.value);
 365
 366     return OvsDoEncapVxlan(vport, curNbl, tunKey, &fwdInfo, layers,
 367                            switchContext, newNbl);
 368 }
 369
 370
 371 /*
 372  *----------------------------------------------------------------------------
 373  * OvsDecapVxlan
 374  *     Decapsulates to tunnel header in 'curNbl' and puts into 'tunKey'.
 375  *----------------------------------------------------------------------------
 376  */
 377 NDIS_STATUS
 378 OvsDecapVxlan(POVS_SWITCH_CONTEXT switchContext,
 379               PNET_BUFFER_LIST curNbl,
 380               OvsIPv4TunnelKey *tunKey,
 381               PNET_BUFFER_LIST *newNbl)
 382 {
 383     PNET_BUFFER curNb;
 384     PMDL curMdl;
 385     EthHdr *ethHdr;
 386     IPHdr *ipHdr;
 387     UDPHdr *udpHdr;
 388     VXLANHdr *vxlanHdr;
 389     UINT32 tunnelSize, packetLength;
 390     PUINT8 bufferStart;
 391     NDIS_STATUS status;
 392     OVS_PACKET_HDR_INFO layers = { 0 };
 393
 394     status = OvsExtractLayers(curNbl, &layers);
 395     if (status != NDIS_STATUS_SUCCESS) {
 396         return status;
 397     }
 398
 399     /* Check the length of the UDP payload */
 400     curNb = NET_BUFFER_LIST_FIRST_NB(curNbl);
 401     packetLength = NET_BUFFER_DATA_LENGTH(curNb);
 402     tunnelSize = OvsGetVxlanTunHdrSizeFromLayers(&layers);
 403     if (packetLength < tunnelSize) {
 404         return NDIS_STATUS_INVALID_LENGTH;
 405     }
 406
 407     /*
 408      * Create a copy of the NBL so that we have all the headers in one MDL.
 409      */
 410     *newNbl = OvsPartialCopyNBL(switchContext, curNbl,
 411                                 tunnelSize, 0,
 412                                 TRUE /*copy NBL info */);
 413
 414     if (*newNbl == NULL) {
 415         return NDIS_STATUS_RESOURCES;
 416     }
 417
 418     /* XXX: Handle VLAN header. */
 419     curNbl = *newNbl;
 420     curNb = NET_BUFFER_LIST_FIRST_NB(curNbl);
 421     curMdl = NET_BUFFER_CURRENT_MDL(curNb);
 422     bufferStart = (PUINT8)OvsGetMdlWithLowPriority(curMdl)
 423         + NET_BUFFER_CURRENT_MDL_OFFSET(curNb);
 424     if (!bufferStart) {
 425         status = NDIS_STATUS_RESOURCES;
 426         goto dropNbl;
 427     }
 428
 429     ethHdr = (EthHdr *)bufferStart;
 430     /* XXX: Handle IP options. */
 431     ipHdr = (IPHdr *)(bufferStart + layers.l3Offset);
 432     tunKey->src = ipHdr->saddr;
 433     tunKey->dst = ipHdr->daddr;
 434     tunKey->tos = ipHdr->tos;
 435     tunKey->ttl = ipHdr->ttl;
 436     tunKey->pad = 0;
 437     udpHdr = (UDPHdr *)(bufferStart + layers.l4Offset);
 438
 439     /* Validate if NIC has indicated checksum failure. */
 440     status = OvsValidateUDPChecksum(curNbl, udpHdr->check == 0);
 441     if (status != NDIS_STATUS_SUCCESS) {
 442         goto dropNbl;
 443     }
 444
 445     /* Calculate and verify UDP checksum if NIC didn't do it. */
 446     if (udpHdr->check != 0) {
 447         tunKey->flags |= OVS_TNL_F_CSUM;
 448         status = OvsCalculateUDPChecksum(curNbl, curNb, ipHdr, udpHdr,
 449                                          packetLength, &layers);
 450         if (status != NDIS_STATUS_SUCCESS) {
 451             goto dropNbl;
 452         }
 453     }
 454
 455     vxlanHdr = (VXLANHdr *)((PCHAR)udpHdr + sizeof *udpHdr);
 456     if (vxlanHdr->instanceID) {
 457         tunKey->flags |= OVS_TNL_F_KEY;
 458         tunKey->tunnelId = VXLAN_VNI_TO_TUNNELID(vxlanHdr->vxlanID);
 459     } else {
 460         tunKey->flags &= ~OVS_TNL_F_KEY;
 461         tunKey->tunnelId = 0;
 462     }
 463
 464     /* Clear out the receive flag for the inner packet. */
 465     NET_BUFFER_LIST_INFO(curNbl, TcpIpChecksumNetBufferListInfo) = 0;
 466     NdisAdvanceNetBufferDataStart(curNb, tunnelSize, FALSE, NULL);
 467     return NDIS_STATUS_SUCCESS;
 468
 469 dropNbl:
 470     OvsCompleteNBL(switchContext, *newNbl, TRUE);
 471     *newNbl = NULL;
 472     return status;
 473 }
 474
 475
 476 NDIS_STATUS
 477 OvsSlowPathDecapVxlan(const PNET_BUFFER_LIST packet,
 478                    OvsIPv4TunnelKey *tunnelKey)
 479 {
 480     NDIS_STATUS status = NDIS_STATUS_FAILURE;
 481     UDPHdr udpStorage;
 482     const UDPHdr *udp;
 483     VXLANHdr *VxlanHeader;
 484     VXLANHdr  VxlanHeaderBuffer;
 485     struct IPHdr ip_storage;
 486     const struct IPHdr *nh;
 487     OVS_PACKET_HDR_INFO layers;
 488
 489     layers.value = 0;
 490
 491     do {
 492         nh = OvsGetIp(packet, layers.l3Offset, &ip_storage);
 493         if (nh) {
 494             layers.l4Offset = layers.l3Offset + nh->ihl * 4;
 495         } else {
 496            status = NDIS_STATUS_INVALID_PACKET;
 497            break;
 498         }
 499
 500         /* make sure it's a VXLAN packet */
 501         udp = OvsGetUdp(packet, layers.l4Offset, &udpStorage);
 502         if (udp) {
 503             layers.l7Offset = layers.l4Offset + sizeof *udp;
 504             if (udp->check != 0) {
 505                 tunnelKey->flags |= OVS_TNL_F_CSUM;
 506             }
 507         } else {
 508             break;
 509         }
 510
 511         VxlanHeader = (VXLANHdr *)OvsGetPacketBytes(packet,
 512                                                     sizeof(*VxlanHeader),
 513                                                     layers.l7Offset,
 514                                                     &VxlanHeaderBuffer);
 515
 516         if (VxlanHeader) {
 517             tunnelKey->src = nh->saddr;
 518             tunnelKey->dst = nh->daddr;
 519             tunnelKey->ttl = nh->ttl;
 520             tunnelKey->tos = nh->tos;
 521             if (VxlanHeader->instanceID) {
 522                 tunnelKey->flags |= OVS_TNL_F_KEY;
 523                 tunnelKey->tunnelId = VXLAN_VNI_TO_TUNNELID(VxlanHeader->vxlanID);
 524             } else {
 525                 tunnelKey->flags &= ~OVS_TNL_F_KEY;
 526                 tunnelKey->tunnelId = 0;
 527             }
 528         } else {
 529             break;
 530         }
 531         status = NDIS_STATUS_SUCCESS;
 532
 533     } while(FALSE);
 534
 535     return status;
 536 }
 537
 538 #pragma warning( pop )