datapath-windows/ovsext/User.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 /*
  18  * OvsUser.c
  19  *      Manage packet queue for packet miss for userAction.
  20  */
  21
  22
  23 #include "precomp.h"
  24
  25 #include "Datapath.h"
  26 #include "Switch.h"
  27 #include "Vport.h"
  28 #include "Event.h"
  29 #include "User.h"
  30 #include "PacketIO.h"
  31 #include "Checksum.h"
  32 #include "NetProto.h"
  33 #include "Flow.h"
  34 #include "TunnelIntf.h"
  35
  36 #ifdef OVS_DBG_MOD
  37 #undef OVS_DBG_MOD
  38 #endif
  39 #define OVS_DBG_MOD OVS_DBG_USER
  40 #include "Debug.h"
  41
  42 OVS_USER_PACKET_QUEUE ovsPacketQueues[OVS_MAX_NUM_PACKET_QUEUES];
  43
  44 POVS_PACKET_QUEUE_ELEM OvsGetNextPacket(POVS_OPEN_INSTANCE instance);
  45 extern PNDIS_SPIN_LOCK gOvsCtrlLock;
  46 extern POVS_SWITCH_CONTEXT gOvsSwitchContext;
  47 OVS_USER_STATS ovsUserStats;
  48
  49 static VOID _MapNlAttrToOvsPktExec(PNL_ATTR *nlAttrs, PNL_ATTR *keyAttrs,
  50                                    OvsPacketExecute  *execute);
  51 extern NL_POLICY nlFlowKeyPolicy[];
  52
  53 NTSTATUS
  54 OvsUserInit()
  55 {
  56     UINT32 i;
  57     POVS_USER_PACKET_QUEUE queue;
  58     for (i = 0; i < OVS_MAX_NUM_PACKET_QUEUES; i++) {
  59         queue = &ovsPacketQueues[i];
  60         RtlZeroMemory(queue, sizeof (*queue));
  61         InitializeListHead(&queue->packetList);
  62         NdisAllocateSpinLock(&queue->queueLock);
  63     }
  64     return STATUS_SUCCESS;
  65 }
  66
  67 VOID
  68 OvsUserCleanup()
  69 {
  70     UINT32 i;
  71     POVS_USER_PACKET_QUEUE queue;
  72     for (i = 0; i < OVS_MAX_NUM_PACKET_QUEUES; i++) {
  73         queue = &ovsPacketQueues[i];
  74         ASSERT(IsListEmpty(&queue->packetList));
  75         ASSERT(queue->instance == NULL);
  76         ASSERT(queue->pendingIrp == NULL);
  77         NdisFreeSpinLock(&queue->queueLock);
  78     }
  79 }
  80
  81 static VOID
  82 OvsPurgePacketQueue(POVS_USER_PACKET_QUEUE queue,
  83                     POVS_OPEN_INSTANCE instance)
  84 {
  85     PLIST_ENTRY link, next;
  86     LIST_ENTRY tmp;
  87     POVS_PACKET_QUEUE_ELEM elem;
  88
  89     InitializeListHead(&tmp);
  90     NdisAcquireSpinLock(&queue->queueLock);
  91     if (queue->instance != instance) {
  92         NdisReleaseSpinLock(&queue->queueLock);
  93         return;
  94     }
  95
  96     if (queue->numPackets) {
  97         OvsAppendList(&tmp, &queue->packetList);
  98         queue->numPackets = 0;
  99     }
 100     NdisReleaseSpinLock(&queue->queueLock);
 101     LIST_FORALL_SAFE(&tmp, link, next) {
 102         RemoveEntryList(link);
 103         elem = CONTAINING_RECORD(link, OVS_PACKET_QUEUE_ELEM, link);
 104         OvsFreeMemory(elem);
 105     }
 106 }
 107
 108
 109 VOID
 110 OvsCleanupPacketQueue(POVS_OPEN_INSTANCE instance)
 111 {
 112     POVS_USER_PACKET_QUEUE queue;
 113     POVS_PACKET_QUEUE_ELEM elem;
 114     PLIST_ENTRY link, next;
 115     LIST_ENTRY tmp;
 116     PIRP irp = NULL;
 117
 118     InitializeListHead(&tmp);
 119     queue = (POVS_USER_PACKET_QUEUE)instance->packetQueue;
 120     if (queue) {
 121         PDRIVER_CANCEL cancelRoutine;
 122         NdisAcquireSpinLock(&queue->queueLock);
 123         if (queue->instance != instance) {
 124             NdisReleaseSpinLock(&queue->queueLock);
 125             return;
 126         }
 127
 128         if (queue->numPackets) {
 129             OvsAppendList(&tmp, &queue->packetList);
 130             queue->numPackets = 0;
 131         }
 132         queue->instance = NULL;
 133         queue->queueId = OVS_MAX_NUM_PACKET_QUEUES;
 134         instance->packetQueue = NULL;
 135         irp = queue->pendingIrp;
 136         queue->pendingIrp = NULL;
 137         if (irp) {
 138             cancelRoutine = IoSetCancelRoutine(irp, NULL);
 139             if (cancelRoutine == NULL) {
 140                 irp = NULL;
 141             }
 142         }
 143         NdisReleaseSpinLock(&queue->queueLock);
 144     }
 145     LIST_FORALL_SAFE(&tmp, link, next) {
 146         RemoveEntryList(link);
 147         elem = CONTAINING_RECORD(link, OVS_PACKET_QUEUE_ELEM, link);
 148         OvsFreeMemory(elem);
 149     }
 150     if (irp) {
 151         OvsCompleteIrpRequest(irp, 0, STATUS_SUCCESS);
 152     }
 153 }
 154
 155 NTSTATUS
 156 OvsSubscribeDpIoctl(PFILE_OBJECT fileObject,
 157                     PVOID inputBuffer,
 158                     UINT32 inputLength)
 159 {
 160     POVS_OPEN_INSTANCE instance = (POVS_OPEN_INSTANCE)fileObject->FsContext;
 161     UINT32 queueId;
 162     POVS_USER_PACKET_QUEUE queue;
 163     if (inputLength < sizeof (UINT32)) {
 164         return STATUS_INVALID_PARAMETER;
 165     }
 166     queueId = *(UINT32 *)inputBuffer;
 167     if (instance->packetQueue && queueId >= OVS_MAX_NUM_PACKET_QUEUES) {
 168         /*
 169          * unsubscribe
 170          */
 171         OvsCleanupPacketQueue(instance);
 172     } else if (instance->packetQueue == NULL &&
 173                queueId < OVS_MAX_NUM_PACKET_QUEUES) {
 174         queue = &ovsPacketQueues[queueId];
 175         NdisAcquireSpinLock(&queue->queueLock);
 176         if (ovsPacketQueues[queueId].instance) {
 177              if (ovsPacketQueues[queueId].instance != instance) {
 178                  NdisReleaseSpinLock(&queue->queueLock);
 179                  return STATUS_INSUFFICIENT_RESOURCES;
 180              } else {
 181                  NdisReleaseSpinLock(&queue->queueLock);
 182                  return STATUS_SUCCESS;
 183              }
 184         }
 185         queue->queueId = queueId;
 186         queue->instance = instance;
 187         instance->packetQueue = queue;
 188         ASSERT(IsListEmpty(&queue->packetList));
 189         NdisReleaseSpinLock(&queue->queueLock);
 190     } else {
 191         return STATUS_INVALID_PARAMETER;
 192     }
 193     return STATUS_SUCCESS;
 194 }
 195
 196
 197 NTSTATUS
 198 OvsReadDpIoctl(PFILE_OBJECT fileObject,
 199                PVOID outputBuffer,
 200                UINT32 outputLength,
 201                UINT32 *replyLen)
 202 {
 203     POVS_OPEN_INSTANCE instance = (POVS_OPEN_INSTANCE)fileObject->FsContext;
 204     POVS_PACKET_QUEUE_ELEM elem;
 205     UINT32 len;
 206
 207 #define TCP_CSUM_OFFSET  16
 208 #define UDP_CSUM_OFFSET  6
 209     ASSERT(instance);
 210
 211     if (instance->packetQueue == NULL) {
 212         return STATUS_INVALID_PARAMETER;
 213     }
 214     if (outputLength < (sizeof (OVS_PACKET_INFO) + OVS_MIN_PACKET_SIZE)) {
 215         return STATUS_BUFFER_TOO_SMALL;
 216     }
 217
 218     elem = OvsGetNextPacket(instance);
 219     if (elem) {
 220         /*
 221          * XXX revisit this later
 222          */
 223         len = elem->packet.totalLen > outputLength ? outputLength :
 224                  elem->packet.totalLen;
 225
 226         if ((elem->hdrInfo.tcpCsumNeeded || elem->hdrInfo.udpCsumNeeded) &&
 227             len == elem->packet.totalLen) {
 228             UINT16 sum, *ptr;
 229             UINT16 size = (UINT16)(elem->packet.payload - elem->packet.data +
 230                                   elem->hdrInfo.l4Offset);
 231             RtlCopyMemory(outputBuffer, &elem->packet.data, size);
 232             ASSERT(len - size >= elem->hdrInfo.l4PayLoad);
 233             sum = CopyAndCalculateChecksum((UINT8 *)outputBuffer + size,
 234                                            (UINT8 *)&elem->packet.data + size,
 235                                            elem->hdrInfo.l4PayLoad, 0);
 236             ptr =(UINT16 *)((UINT8 *)outputBuffer + size +
 237                             (elem->hdrInfo.tcpCsumNeeded ?
 238                              TCP_CSUM_OFFSET : UDP_CSUM_OFFSET));
 239             *ptr = sum;
 240             ovsUserStats.l4Csum++;
 241         } else {
 242             RtlCopyMemory(outputBuffer, &elem->packet, len);
 243         }
 244
 245         *replyLen = len;
 246         OvsFreeMemory(elem);
 247     }
 248     return STATUS_SUCCESS;
 249 }
 250
 251 /* Helper function to allocate a Forwarding Context for an NBL */
 252 NTSTATUS
 253 OvsAllocateForwardingContextForNBL(POVS_SWITCH_CONTEXT switchContext,
 254                                    PNET_BUFFER_LIST nbl)
 255 {
 256     return switchContext->NdisSwitchHandlers.
 257         AllocateNetBufferListForwardingContext(
 258             switchContext->NdisSwitchContext, nbl);
 259 }
 260
 261 /*
 262  * --------------------------------------------------------------------------
 263  * This function allocates all the stuff necessary for creating an NBL from the
 264  * input buffer of specified length, namely, a nonpaged data buffer of size
 265  * length, an MDL from it, and a NB and NBL from it. It does not allocate an NBL
 266  * context yet. It also copies data from the specified buffer to the NBL.
 267  * --------------------------------------------------------------------------
 268  */
 269 PNET_BUFFER_LIST
 270 OvsAllocateNBLForUserBuffer(POVS_SWITCH_CONTEXT switchContext,
 271                             PVOID userBuffer,
 272                             ULONG length)
 273 {
 274     UINT8 *data = NULL;
 275     PNET_BUFFER_LIST nbl = NULL;
 276     PNET_BUFFER nb;
 277     PMDL mdl;
 278
 279     if (length > OVS_DEFAULT_DATA_SIZE) {
 280         nbl = OvsAllocateVariableSizeNBL(switchContext, length,
 281                                          OVS_DEFAULT_HEADROOM_SIZE);
 282
 283     } else {
 284         nbl = OvsAllocateFixSizeNBL(switchContext, length,
 285                                     OVS_DEFAULT_HEADROOM_SIZE);
 286     }
 287     if (nbl == NULL) {
 288         return NULL;
 289     }
 290
 291     nb = NET_BUFFER_LIST_FIRST_NB(nbl);
 292     mdl = NET_BUFFER_CURRENT_MDL(nb);
 293     data = (PUINT8)MmGetSystemAddressForMdlSafe(mdl, LowPagePriority) +
 294                     NET_BUFFER_CURRENT_MDL_OFFSET(nb);
 295     if (!data) {
 296         OvsCompleteNBL(switchContext, nbl, TRUE);
 297         return NULL;
 298     }
 299
 300     NdisMoveMemory(data, userBuffer, length);
 301
 302     return nbl;
 303 }
 304
 305 /*
 306  *----------------------------------------------------------------------------
 307  *  OvsNlExecuteCmdHandler --
 308  *    Handler for OVS_PACKET_CMD_EXECUTE command.
 309  *----------------------------------------------------------------------------
 310  */
 311 NTSTATUS
 312 OvsNlExecuteCmdHandler(POVS_USER_PARAMS_CONTEXT usrParamsCtx,
 313                        UINT32 *replyLen)
 314 {
 315     NTSTATUS status = STATUS_SUCCESS;
 316     POVS_MESSAGE msgIn = (POVS_MESSAGE)usrParamsCtx->inputBuffer;
 317     POVS_MESSAGE msgOut = (POVS_MESSAGE)usrParamsCtx->outputBuffer;
 318     PNL_MSG_HDR nlMsgHdr = &(msgIn->nlMsg);
 319     PGENL_MSG_HDR genlMsgHdr = &(msgIn->genlMsg);
 320     POVS_HDR ovsHdr = &(msgIn->ovsHdr);
 321
 322     PNL_ATTR nlAttrs[__OVS_PACKET_ATTR_MAX];
 323     PNL_ATTR keyAttrs[__OVS_KEY_ATTR_MAX] = {NULL};
 324
 325     UINT32 attrOffset = NLMSG_HDRLEN + GENL_HDRLEN + OVS_HDRLEN;
 326     UINT32 keyAttrOffset = 0;
 327     OvsPacketExecute execute;
 328     NL_ERROR nlError = NL_ERROR_SUCCESS;
 329     NL_BUFFER nlBuf;
 330
 331     static const NL_POLICY nlPktExecPolicy[] = {
 332         [OVS_PACKET_ATTR_PACKET] = {.type = NL_A_UNSPEC, .optional = FALSE},
 333         [OVS_PACKET_ATTR_KEY] = {.type = NL_A_UNSPEC, .optional = FALSE},
 334         [OVS_PACKET_ATTR_ACTIONS] = {.type = NL_A_UNSPEC, .optional = FALSE},
 335         [OVS_PACKET_ATTR_USERDATA] = {.type = NL_A_UNSPEC, .optional = TRUE},
 336         [OVS_PACKET_ATTR_EGRESS_TUN_KEY] = {.type = NL_A_UNSPEC,
 337                                             .optional = TRUE}
 338     };
 339
 340     RtlZeroMemory(&execute, sizeof(OvsPacketExecute));
 341
 342     /* Get all the top level Flow attributes */
 343     if ((NlAttrParse(nlMsgHdr, attrOffset, NlMsgAttrsLen(nlMsgHdr),
 344                      nlPktExecPolicy, nlAttrs, ARRAY_SIZE(nlAttrs)))
 345                      != TRUE) {
 346         OVS_LOG_ERROR("Attr Parsing failed for msg: %p",
 347                        nlMsgHdr);
 348         status = STATUS_UNSUCCESSFUL;
 349         goto done;
 350     }
 351
 352     keyAttrOffset = (UINT32)((PCHAR)nlAttrs[OVS_PACKET_ATTR_KEY] -
 353                     (PCHAR)nlMsgHdr);
 354
 355     /* Get flow keys attributes */
 356     if ((NlAttrParseNested(nlMsgHdr, keyAttrOffset,
 357                            NlAttrLen(nlAttrs[OVS_PACKET_ATTR_KEY]),
 358                            nlFlowKeyPolicy, keyAttrs,
 359                            ARRAY_SIZE(keyAttrs))) != TRUE) {
 360         OVS_LOG_ERROR("Key Attr Parsing failed for msg: %p", nlMsgHdr);
 361         status = STATUS_UNSUCCESSFUL;
 362         goto done;
 363     }
 364
 365     execute.dpNo = ovsHdr->dp_ifindex;
 366
 367     _MapNlAttrToOvsPktExec(nlAttrs, keyAttrs, &execute);
 368
 369     status = OvsExecuteDpIoctl(&execute);
 370
 371     /* Default reply that we want to send */
 372     if (status == STATUS_SUCCESS) {
 373         NlBufInit(&nlBuf, usrParamsCtx->outputBuffer,
 374                   usrParamsCtx->outputLength);
 375
 376         /* Prepare nl Msg headers */
 377         status = NlFillOvsMsg(&nlBuf, nlMsgHdr->nlmsgType, 0,
 378                  nlMsgHdr->nlmsgSeq, nlMsgHdr->nlmsgPid,
 379                  genlMsgHdr->cmd, OVS_PACKET_VERSION,
 380                  ovsHdr->dp_ifindex);
 381
 382         if (status == STATUS_SUCCESS) {
 383             *replyLen = msgOut->nlMsg.nlmsgLen;
 384         }
 385     } else {
 386         /* Map NTSTATUS to NL_ERROR */
 387         nlError = NlMapStatusToNlErr(status);
 388
 389         /* As of now there are no transactional errors in the implementation.
 390          * Once we have them then we need to map status to correct
 391          * nlError value, so that below mentioned code gets hit. */
 392         if ((nlError != NL_ERROR_SUCCESS) &&
 393             (usrParamsCtx->outputBuffer)) {
 394
 395             POVS_MESSAGE_ERROR msgError = (POVS_MESSAGE_ERROR)
 396                                            usrParamsCtx->outputBuffer;
 397             BuildErrorMsg(msgIn, msgError, nlError);
 398             *replyLen = msgError->nlMsg.nlmsgLen;
 399             status = STATUS_SUCCESS;
 400             goto done;
 401         }
 402     }
 403
 404 done:
 405     return status;
 406 }
 407
 408 /*
 409  *----------------------------------------------------------------------------
 410  *  _MapNlAttrToOvsPktExec --
 411  *    Maps input Netlink attributes to OvsPacketExecute.
 412  *----------------------------------------------------------------------------
 413  */
 414 static VOID
 415 _MapNlAttrToOvsPktExec(PNL_ATTR *nlAttrs, PNL_ATTR *keyAttrs,
 416                        OvsPacketExecute *execute)
 417 {
 418     execute->packetBuf = NlAttrGet(nlAttrs[OVS_PACKET_ATTR_PACKET]);
 419     execute->packetLen = NlAttrGetSize(nlAttrs[OVS_PACKET_ATTR_PACKET]);
 420
 421     execute->actions = NlAttrGet(nlAttrs[OVS_PACKET_ATTR_ACTIONS]);
 422     execute->actionsLen = NlAttrGetSize(nlAttrs[OVS_PACKET_ATTR_ACTIONS]);
 423
 424     execute->inPort = NlAttrGetU32(keyAttrs[OVS_KEY_ATTR_IN_PORT]);
 425 }
 426
 427 NTSTATUS
 428 OvsExecuteDpIoctl(OvsPacketExecute *execute)
 429 {
 430     NTSTATUS                    status = STATUS_SUCCESS;
 431     NTSTATUS                    ndisStatus;
 432     LOCK_STATE_EX               lockState;
 433     PNET_BUFFER_LIST pNbl;
 434     PNL_ATTR actions;
 435     PNDIS_SWITCH_FORWARDING_DETAIL_NET_BUFFER_LIST_INFO fwdDetail;
 436     OvsFlowKey key;
 437     OVS_PACKET_HDR_INFO layers;
 438     POVS_VPORT_ENTRY vport;
 439
 440     NdisAcquireSpinLock(gOvsCtrlLock);
 441     if (gOvsSwitchContext == NULL) {
 442         status = STATUS_INVALID_PARAMETER;
 443         goto unlock;
 444     }
 445
 446     if (execute->packetLen == 0) {
 447         status = STATUS_INVALID_PARAMETER;
 448         goto unlock;
 449     }
 450
 451     actions = execute->actions;
 452
 453     ASSERT(actions);
 454
 455     /*
 456      * Allocate the NBL, copy the data from the userspace buffer. Allocate
 457      * also, the forwarding context for the packet.
 458      */
 459     pNbl = OvsAllocateNBLForUserBuffer(gOvsSwitchContext, execute->packetBuf,
 460                                        execute->packetLen);
 461     if (pNbl == NULL) {
 462         status = STATUS_NO_MEMORY;
 463         goto unlock;
 464     }
 465
 466     fwdDetail = NET_BUFFER_LIST_SWITCH_FORWARDING_DETAIL(pNbl);
 467     vport = OvsFindVportByPortNo(gOvsSwitchContext, execute->inPort);
 468     if (vport) {
 469         fwdDetail->SourcePortId = vport->portId;
 470         fwdDetail->SourceNicIndex = vport->nicIndex;
 471     } else {
 472         fwdDetail->SourcePortId = NDIS_SWITCH_DEFAULT_PORT_ID;
 473         fwdDetail->SourceNicIndex = 0;
 474     }
 475     // XXX: Figure out if any of the other members of fwdDetail need to be set.
 476
 477     ndisStatus = OvsExtractFlow(pNbl, fwdDetail->SourcePortId, &key, &layers,
 478                               NULL);
 479     if (ndisStatus == NDIS_STATUS_SUCCESS) {
 480         ASSERT(KeGetCurrentIrql() == DISPATCH_LEVEL);
 481         NdisAcquireRWLockRead(gOvsSwitchContext->dispatchLock, &lockState,
 482                               NDIS_RWL_AT_DISPATCH_LEVEL);
 483         ndisStatus = OvsActionsExecute(gOvsSwitchContext, NULL, pNbl,
 484                                        vport ? vport->portNo :
 485                                                OVS_DEFAULT_PORT_NO,
 486                                        NDIS_SEND_FLAGS_SWITCH_DESTINATION_GROUP,
 487                                        &key, NULL, &layers, actions,
 488                                        execute->actionsLen);
 489         pNbl = NULL;
 490         NdisReleaseRWLock(gOvsSwitchContext->dispatchLock, &lockState);
 491     }
 492     if (ndisStatus != NDIS_STATUS_SUCCESS) {
 493         if (ndisStatus == NDIS_STATUS_NOT_SUPPORTED) {
 494             status = STATUS_NOT_SUPPORTED;
 495         } else {
 496             status = STATUS_UNSUCCESSFUL;
 497         }
 498     }
 499
 500     if (pNbl) {
 501         OvsCompleteNBL(gOvsSwitchContext, pNbl, TRUE);
 502     }
 503 unlock:
 504     NdisReleaseSpinLock(gOvsCtrlLock);
 505     return status;
 506 }
 507
 508
 509 NTSTATUS
 510 OvsPurgeDpIoctl(PFILE_OBJECT fileObject)
 511 {
 512     POVS_OPEN_INSTANCE instance = (POVS_OPEN_INSTANCE)fileObject->FsContext;
 513     POVS_USER_PACKET_QUEUE queue = (POVS_USER_PACKET_QUEUE)instance->packetQueue;
 514
 515     if (queue == NULL) {
 516         return STATUS_INVALID_PARAMETER;
 517     }
 518     OvsPurgePacketQueue(queue, instance);
 519     return STATUS_SUCCESS;
 520 }
 521
 522 VOID
 523 OvsCancelIrpDatapath(PDEVICE_OBJECT deviceObject,
 524                      PIRP irp)
 525 {
 526     PIO_STACK_LOCATION irpSp;
 527     PFILE_OBJECT fileObject;
 528     POVS_OPEN_INSTANCE instance;
 529     POVS_USER_PACKET_QUEUE queue = NULL;
 530
 531     UNREFERENCED_PARAMETER(deviceObject);
 532
 533     IoReleaseCancelSpinLock(irp->CancelIrql);
 534     irpSp = IoGetCurrentIrpStackLocation(irp);
 535     fileObject = irpSp->FileObject;
 536
 537     if (fileObject == NULL) {
 538         goto done;
 539     }
 540     NdisAcquireSpinLock(gOvsCtrlLock);
 541     instance = (POVS_OPEN_INSTANCE)fileObject->FsContext;
 542     if (instance) {
 543         queue = instance->packetQueue;
 544     }
 545     if (instance == NULL || queue == NULL) {
 546         NdisReleaseSpinLock(gOvsCtrlLock);
 547         goto done;
 548     }
 549     NdisReleaseSpinLock(gOvsCtrlLock);
 550     NdisAcquireSpinLock(&queue->queueLock);
 551     if (queue->pendingIrp == irp) {
 552         queue->pendingIrp = NULL;
 553     }
 554     NdisReleaseSpinLock(&queue->queueLock);
 555 done:
 556     OvsCompleteIrpRequest(irp, 0, STATUS_CANCELLED);
 557 }
 558
 559
 560 NTSTATUS
 561 OvsWaitDpIoctl(PIRP irp, PFILE_OBJECT fileObject)
 562 {
 563     POVS_OPEN_INSTANCE instance = (POVS_OPEN_INSTANCE)fileObject->FsContext;
 564     POVS_USER_PACKET_QUEUE queue =
 565                (POVS_USER_PACKET_QUEUE)instance->packetQueue;
 566     NTSTATUS status = STATUS_SUCCESS;
 567     BOOLEAN cancelled = FALSE;
 568
 569     if (queue == NULL) {
 570         return STATUS_INVALID_PARAMETER;
 571     }
 572     NdisAcquireSpinLock(&queue->queueLock);
 573     if (queue->instance != instance) {
 574         NdisReleaseSpinLock(&queue->queueLock);
 575         return STATUS_INVALID_PARAMETER;
 576     }
 577     if (queue->pendingIrp) {
 578         NdisReleaseSpinLock(&queue->queueLock);
 579         return STATUS_DEVICE_BUSY;
 580     }
 581     if (queue->numPackets == 0) {
 582         PDRIVER_CANCEL cancelRoutine;
 583         IoMarkIrpPending(irp);
 584         IoSetCancelRoutine(irp, OvsCancelIrpDatapath);
 585         if (irp->Cancel) {
 586             cancelRoutine = IoSetCancelRoutine(irp, NULL);
 587             if (cancelRoutine) {
 588                 cancelled = TRUE;
 589             }
 590         } else {
 591             queue->pendingIrp = irp;
 592         }
 593         status = STATUS_PENDING;
 594     }
 595     NdisReleaseSpinLock(&queue->queueLock);
 596     if (cancelled) {
 597         OvsCompleteIrpRequest(irp, 0, STATUS_CANCELLED);
 598         OVS_LOG_INFO("Datapath IRP cancelled: %p", irp);
 599     }
 600     return status;
 601 }
 602
 603
 604 POVS_PACKET_QUEUE_ELEM
 605 OvsGetNextPacket(POVS_OPEN_INSTANCE instance)
 606 {
 607     POVS_USER_PACKET_QUEUE queue;
 608     PLIST_ENTRY link;
 609     queue = (POVS_USER_PACKET_QUEUE)instance->packetQueue;
 610     if (queue == NULL) {
 611         return NULL;
 612     }
 613     NdisAcquireSpinLock(&queue->queueLock);
 614     if (queue->instance != instance || queue->numPackets == 0) {
 615         NdisReleaseSpinLock(&queue->queueLock);
 616         return NULL;
 617     }
 618     link = RemoveHeadList(&queue->packetList);
 619     queue->numPackets--;
 620     NdisReleaseSpinLock(&queue->queueLock);
 621     return CONTAINING_RECORD(link, OVS_PACKET_QUEUE_ELEM, link);
 622 }
 623
 624
 625 POVS_USER_PACKET_QUEUE
 626 OvsGetQueue(UINT32 queueId)
 627 {
 628     POVS_USER_PACKET_QUEUE queue;
 629     if (queueId >= OVS_MAX_NUM_PACKET_QUEUES) {
 630         return NULL;
 631     }
 632     queue = &ovsPacketQueues[queueId];
 633     return queue->instance != NULL ? queue : NULL;
 634 }
 635
 636 VOID
 637 OvsQueuePackets(UINT32 queueId,
 638                 PLIST_ENTRY packetList,
 639                 UINT32 numElems)
 640 {
 641     POVS_USER_PACKET_QUEUE queue = OvsGetQueue(queueId);
 642     POVS_PACKET_QUEUE_ELEM elem;
 643     PIRP irp = NULL;
 644     PLIST_ENTRY  link;
 645     UINT32 num = 0;
 646
 647     OVS_LOG_LOUD("Enter: queueId %u, numELems: %u",
 648                   queueId, numElems);
 649     if (queue == NULL) {
 650         goto cleanup;
 651     }
 652
 653     NdisAcquireSpinLock(&queue->queueLock);
 654     if (queue->instance == NULL) {
 655         NdisReleaseSpinLock(&queue->queueLock);
 656         goto cleanup;
 657     } else {
 658         OvsAppendList(&queue->packetList, packetList);
 659         queue->numPackets += numElems;
 660     }
 661     if (queue->pendingIrp) {
 662         PDRIVER_CANCEL cancelRoutine;
 663         irp = queue->pendingIrp;
 664         queue->pendingIrp = NULL;
 665         cancelRoutine = IoSetCancelRoutine(irp, NULL);
 666         if (cancelRoutine == NULL) {
 667             irp = NULL;
 668         }
 669     }
 670     NdisReleaseSpinLock(&queue->queueLock);
 671     if (irp) {
 672         OvsCompleteIrpRequest(irp, 0, STATUS_SUCCESS);
 673     }
 674
 675 cleanup:
 676     while (!IsListEmpty(packetList)) {
 677         link = RemoveHeadList(packetList);
 678         elem = CONTAINING_RECORD(link, OVS_PACKET_QUEUE_ELEM, link);
 679         OvsFreeMemory(elem);
 680         num++;
 681     }
 682     OVS_LOG_LOUD("Exit: drop %u packets", num);
 683 }
 684
 685
 686 /*
 687  *----------------------------------------------------------------------------
 688  * OvsCreateAndAddPackets --
 689  *
 690  *  Create a packet and forwarded to user space.
 691  *
 692  *  This function would fragment packet if needed, and queue
 693  *  each segment to user space.
 694  *----------------------------------------------------------------------------
 695  */
 696 NTSTATUS
 697 OvsCreateAndAddPackets(PVOID userData,
 698                        UINT32 userDataLen,
 699                        UINT32 cmd,
 700                        UINT32 inPort,
 701                        OvsFlowKey *key,
 702                        PNET_BUFFER_LIST nbl,
 703                        BOOLEAN isRecv,
 704                        POVS_PACKET_HDR_INFO hdrInfo,
 705                        POVS_SWITCH_CONTEXT switchContext,
 706                        LIST_ENTRY *list,
 707                        UINT32 *num)
 708 {
 709     POVS_PACKET_QUEUE_ELEM elem;
 710     PNET_BUFFER_LIST newNbl = NULL;
 711     PNET_BUFFER nb;
 712
 713     if (hdrInfo->isTcp) {
 714         NDIS_TCP_LARGE_SEND_OFFLOAD_NET_BUFFER_LIST_INFO tsoInfo;
 715         UINT32 packetLength;
 716
 717         tsoInfo.Value = NET_BUFFER_LIST_INFO(nbl, TcpLargeSendNetBufferListInfo);
 718         nb = NET_BUFFER_LIST_FIRST_NB(nbl);
 719         packetLength = NET_BUFFER_DATA_LENGTH(nb);
 720
 721         OVS_LOG_TRACE("MSS %u packet len %u",
 722                 tsoInfo.LsoV1Transmit.MSS, packetLength);
 723         if (tsoInfo.LsoV1Transmit.MSS) {
 724             OVS_LOG_TRACE("l4Offset %d", hdrInfo->l4Offset);
 725             newNbl = OvsTcpSegmentNBL(switchContext, nbl, hdrInfo,
 726                     tsoInfo.LsoV1Transmit.MSS , 0);
 727             if (newNbl == NULL) {
 728                 return NDIS_STATUS_FAILURE;
 729             }
 730             nbl = newNbl;
 731         }
 732     }
 733
 734     nb = NET_BUFFER_LIST_FIRST_NB(nbl);
 735     while (nb) {
 736         elem = OvsCreateQueueNlPacket(userData, userDataLen,
 737                                     cmd, inPort, key, nbl, nb,
 738                                     isRecv, hdrInfo);
 739         if (elem) {
 740             InsertTailList(list, &elem->link);
 741             (*num)++;
 742         }
 743         nb = NET_BUFFER_NEXT_NB(nb);
 744     }
 745     if (newNbl) {
 746         OvsCompleteNBL(switchContext, newNbl, TRUE);
 747     }
 748     return NDIS_STATUS_SUCCESS;
 749 }
 750
 751 static __inline UINT32
 752 OvsGetUpcallMsgSize(PVOID userData,
 753                     UINT32 userDataLen,
 754                     OvsIPv4TunnelKey *tunnelKey,
 755                     UINT32 payload)
 756 {
 757     UINT32 size = NLMSG_ALIGN(sizeof(struct ovs_header)) +
 758                   NlAttrSize(payload) +
 759                   NlAttrSize(OvsFlowKeyAttrSize());
 760
 761     /* OVS_PACKET_ATTR_USERDATA */
 762     if (userData) {
 763         size += NlAttrTotalSize(userDataLen);
 764     }
 765     /* OVS_PACKET_ATTR_EGRESS_TUN_KEY */
 766     /* Is it included in the the flwo key attr XXX */
 767     if (tunnelKey) {
 768         size += NlAttrTotalSize(OvsTunKeyAttrSize());
 769     }
 770     return size;
 771 }
 772
 773 /*
 774  *----------------------------------------------------------------------------
 775  * This function completes the IP Header csum. record the L4 payload offset and
 776  * if there is a need to calculate the TCP or UDP csum. The actual csum will be
 777  * caluculated simopultaneossly with the copy of the payload to the destination
 778  * buffer when the packet is read.
 779  *----------------------------------------------------------------------------
 780  */
 781 static VOID
 782 OvsCompletePacketHeader(UINT8 *packet,
 783                         BOOLEAN isRecv,
 784                         NDIS_TCP_IP_CHECKSUM_NET_BUFFER_LIST_INFO csumInfo,
 785                         POVS_PACKET_HDR_INFO hdrInfoIn,
 786                         POVS_PACKET_HDR_INFO hdrInfoOut)
 787 {
 788     if ((isRecv && csumInfo.Receive.IpChecksumValueInvalid) ||
 789         (!isRecv && csumInfo.Transmit.IsIPv4 &&
 790         csumInfo.Transmit.IpHeaderChecksum)) {
 791         PIPV4_HEADER ipHdr = (PIPV4_HEADER)(packet + hdrInfoOut->l3Offset);
 792         ASSERT(hdrInfoIn->isIPv4);
 793         ASSERT(ipHdr->Version == 4);
 794         ipHdr->HeaderChecksum = IPChecksum((UINT8 *)ipHdr,
 795             ipHdr->HeaderLength << 2,
 796             (UINT16)~ipHdr->HeaderChecksum);
 797         ovsUserStats.ipCsum++;
 798     }
 799     ASSERT(hdrInfoIn->tcpCsumNeeded == 0 && hdrInfoOut->udpCsumNeeded == 0);
 800     /*
 801      * calculate TCP/UDP pseudo checksum
 802      */
 803     if (isRecv && csumInfo.Receive.TcpChecksumValueInvalid) {
 804         /*
 805          * Only this case, we need to reclaculate pseudo checksum
 806          * all other cases, it is assumed the pseudo checksum is
 807          * filled already.
 808          *
 809          */
 810         PTCP_HDR tcpHdr = (PTCP_HDR)(packet + hdrInfoIn->l4Offset);
 811         if (hdrInfoIn->isIPv4) {
 812             PIPV4_HEADER ipHdr = (PIPV4_HEADER)(packet + hdrInfoIn->l3Offset);
 813             hdrInfoOut->l4PayLoad = (UINT16)(ntohs(ipHdr->TotalLength) -
 814                                     (ipHdr->HeaderLength << 2));
 815             tcpHdr->th_sum = IPPseudoChecksum((UINT32 *)&ipHdr->SourceAddress,
 816                                          (UINT32 *)&ipHdr->DestinationAddress,
 817                                          IPPROTO_TCP, hdrInfoOut->l4PayLoad);
 818         } else {
 819             PIPV6_HEADER ipv6Hdr = (PIPV6_HEADER)(packet + hdrInfoIn->l3Offset);
 820             hdrInfoOut->l4PayLoad =
 821                 (UINT16)(ntohs(ipv6Hdr->PayloadLength) +
 822                 hdrInfoIn->l3Offset + sizeof(IPV6_HEADER)-
 823                 hdrInfoIn->l4Offset);
 824             ASSERT(hdrInfoIn->isIPv6);
 825             tcpHdr->th_sum =
 826                 IPv6PseudoChecksum((UINT32 *)&ipv6Hdr->SourceAddress,
 827                 (UINT32 *)&ipv6Hdr->DestinationAddress,
 828                 IPPROTO_TCP, hdrInfoOut->l4PayLoad);
 829         }
 830         hdrInfoOut->tcpCsumNeeded = 1;
 831         ovsUserStats.recalTcpCsum++;
 832     } else if (!isRecv) {
 833         if (csumInfo.Transmit.TcpChecksum) {
 834             hdrInfoOut->tcpCsumNeeded = 1;
 835         } else if (csumInfo.Transmit.UdpChecksum) {
 836             hdrInfoOut->udpCsumNeeded = 1;
 837         }
 838         if (hdrInfoOut->tcpCsumNeeded || hdrInfoOut->udpCsumNeeded) {
 839 #ifdef DBG
 840             UINT16 sum, *ptr;
 841             UINT8 proto =
 842                 hdrInfoOut->tcpCsumNeeded ? IPPROTO_TCP : IPPROTO_UDP;
 843 #endif
 844             if (hdrInfoIn->isIPv4) {
 845                 PIPV4_HEADER ipHdr = (PIPV4_HEADER)(packet + hdrInfoIn->l3Offset);
 846                 hdrInfoOut->l4PayLoad = (UINT16)(ntohs(ipHdr->TotalLength) -
 847                     (ipHdr->HeaderLength << 2));
 848 #ifdef DBG
 849                 sum = IPPseudoChecksum((UINT32 *)&ipHdr->SourceAddress,
 850                     (UINT32 *)&ipHdr->DestinationAddress,
 851                     proto, hdrInfoOut->l4PayLoad);
 852 #endif
 853             } else {
 854                 PIPV6_HEADER ipv6Hdr = (PIPV6_HEADER)(packet +
 855                     hdrInfoIn->l3Offset);
 856                 hdrInfoOut->l4PayLoad =
 857                     (UINT16)(ntohs(ipv6Hdr->PayloadLength) +
 858                     hdrInfoIn->l3Offset + sizeof(IPV6_HEADER)-
 859                     hdrInfoIn->l4Offset);
 860                 ASSERT(hdrInfoIn->isIPv6);
 861 #ifdef DBG
 862                 sum = IPv6PseudoChecksum((UINT32 *)&ipv6Hdr->SourceAddress,
 863                     (UINT32 *)&ipv6Hdr->DestinationAddress,
 864                     proto, hdrInfoOut->l4PayLoad);
 865 #endif
 866             }
 867 #ifdef DBG
 868             ptr = (UINT16 *)(packet + hdrInfoIn->l4Offset +
 869                 (hdrInfoOut->tcpCsumNeeded ?
 870             TCP_CSUM_OFFSET : UDP_CSUM_OFFSET));
 871             ASSERT(*ptr == sum);
 872 #endif
 873         }
 874     }
 875 }
 876
 877 static NTSTATUS
 878 OvsGetPid(POVS_VPORT_ENTRY vport, PNET_BUFFER nb, UINT32 *pid)
 879 {
 880     UNREFERENCED_PARAMETER(nb);
 881
 882     /* XXX select a pid from an array of pids using a flow based hash */
 883     *pid = vport->upcallPid;
 884     return STATUS_SUCCESS;
 885 }
 886
 887 /*
 888  *----------------------------------------------------------------------------
 889  * OvsCreateQueueNlPacket --
 890  *
 891  *  Create a packet which will be forwarded to user space.
 892  *
 893  * InputParameter:
 894  *   userData: when cmd is user action, this field contain
 895  *      user action data.
 896  *   userDataLen: as name indicated
 897  *   cmd: either miss or user action
 898  *   inPort: datapath port id from which the packet is received.
 899  *   key: flow Key with a tunnel key if available
 900  *   nbl:  the NET_BUFFER_LIST which contain the packet
 901  *   nb: the packet
 902  *   isRecv: This is used to decide how to interprete the csum info
 903  *   hdrInfo: include hdr info initialized during flow extraction.
 904  *
 905  * Results:
 906  *    NULL if fail to create the packet
 907  *    The packet element otherwise
 908  *----------------------------------------------------------------------------
 909  */
 910 POVS_PACKET_QUEUE_ELEM
 911 OvsCreateQueueNlPacket(PVOID userData,
 912                        UINT32 userDataLen,
 913                        UINT32 cmd,
 914                        UINT32 inPort,
 915                        OvsFlowKey *key,
 916                        PNET_BUFFER_LIST nbl,
 917                        PNET_BUFFER nb,
 918                        BOOLEAN isRecv,
 919                        POVS_PACKET_HDR_INFO hdrInfo)
 920 {
 921 #define VLAN_TAG_SIZE 4
 922     UINT32 allocLen, dataLen, extraLen;
 923     POVS_PACKET_QUEUE_ELEM elem;
 924     UINT8 *src, *dst;
 925     NDIS_TCP_IP_CHECKSUM_NET_BUFFER_LIST_INFO csumInfo;
 926     NDIS_NET_BUFFER_LIST_8021Q_INFO vlanInfo;
 927     OvsIPv4TunnelKey *tunnelKey = (OvsIPv4TunnelKey *)&key->tunKey;
 928     UINT32 pid;
 929     UINT32 nlMsgSize;
 930     NL_BUFFER nlBuf;
 931
 932     /* XXX pass vport in the stack rather than portNo */
 933     POVS_VPORT_ENTRY vport =
 934         OvsFindVportByPortNo(gOvsSwitchContext, inPort);
 935
 936     if (vport == NULL){
 937         /* Should never happen as dispatch lock is held */
 938         ASSERT(vport);
 939         return NULL;
 940     }
 941
 942     if (!OvsGetPid(vport, nb, &pid)) {
 943         /*
 944          * There is no userspace queue created yet, so there is no point for
 945          * creating a new packet to be queued.
 946          */
 947         return NULL;
 948     }
 949
 950     csumInfo.Value = NET_BUFFER_LIST_INFO(nbl, TcpIpChecksumNetBufferListInfo);
 951
 952     if (isRecv && (csumInfo.Receive.TcpChecksumFailed ||
 953                   (csumInfo.Receive.UdpChecksumFailed && !hdrInfo->udpCsumZero) ||
 954                   csumInfo.Receive.IpChecksumFailed)) {
 955         OVS_LOG_INFO("Packet dropped due to checksum failure.");
 956         ovsUserStats.dropDuetoChecksum++;
 957         return NULL;
 958     }
 959
 960     vlanInfo.Value = NET_BUFFER_LIST_INFO(nbl, Ieee8021QNetBufferListInfo);
 961     extraLen = vlanInfo.TagHeader.VlanId ? VLAN_TAG_SIZE : 0;
 962
 963     dataLen = NET_BUFFER_DATA_LENGTH(nb);
 964
 965     if (NlAttrSize(dataLen) > MAXUINT16) {
 966         return NULL;
 967     }
 968
 969     nlMsgSize = OvsGetUpcallMsgSize(userData, userDataLen, tunnelKey,
 970                                     dataLen + extraLen);
 971
 972     allocLen = sizeof (OVS_PACKET_QUEUE_ELEM) + nlMsgSize;
 973     elem = (POVS_PACKET_QUEUE_ELEM)OvsAllocateMemory(allocLen);
 974     if (elem == NULL) {
 975         ovsUserStats.dropDuetoResource++;
 976         return NULL;
 977     }
 978     elem->hdrInfo.value = hdrInfo->value;
 979     elem->packet.totalLen = nlMsgSize;
 980     /* XXX remove queueid */
 981     elem->packet.queue = 0;
 982     /* XXX  no need as the length is already in the NL attrib */
 983     elem->packet.userDataLen = userDataLen;
 984     elem->packet.inPort = inPort;
 985     elem->packet.cmd = cmd;
 986     if (cmd == (UINT32)OVS_PACKET_CMD_MISS) {
 987         ovsUserStats.miss++;
 988     } else if (cmd == (UINT32)OVS_PACKET_CMD_ACTION) {
 989         ovsUserStats.action++;
 990     } else {
 991         ASSERT(FALSE);
 992         goto fail;
 993     }
 994     /* XXX Should we have both packetLen and TotalLen*/
 995     elem->packet.packetLen = dataLen + extraLen;
 996
 997     NlBufInit(&nlBuf, (PCHAR)elem->packet.data, nlMsgSize);
 998
 999     /*
1000      * Initialize the OVS header
1001      * Since we are pre allocating memory for the NL buffer
1002      * the attribute settings should not fail
1003      */
1004     if (NlFillOvsMsg(&nlBuf, OVS_WIN_NL_PACKET_FAMILY_ID, 0,
1005                       0, pid, (UINT8)cmd, OVS_PACKET_VERSION,
1006                       gOvsSwitchContext->dpNo) != STATUS_SUCCESS) {
1007         goto fail;
1008     }
1009
1010     if (MapFlowKeyToNlKey(&nlBuf, key, OVS_PACKET_ATTR_KEY,
1011                           OVS_KEY_ATTR_TUNNEL) != STATUS_SUCCESS) {
1012         goto fail;
1013     }
1014
1015     /* XXX must send OVS_PACKET_ATTR_EGRESS_TUN_KEY if set by vswtchd */
1016     if (userData){
1017         if (!NlMsgPutTailUnspec(&nlBuf, OVS_PACKET_ATTR_USERDATA,
1018                                 userData, (UINT16)userDataLen)) {
1019             goto fail;
1020         }
1021     }
1022
1023     /*
1024      * Make space for the payload to be copied and set the attribute
1025      * XXX Uninit set initilizes the buffer with xero, we don't actually need
1026      * that the payload to be initailized
1027      */
1028     dst = (UINT8 *)NlMsgPutTailUnspecUninit(&nlBuf, OVS_PACKET_ATTR_PACKET,
1029                                             (UINT16)(dataLen + extraLen));
1030     if (!dst) {
1031         goto fail;
1032     }
1033
1034     /* Store the payload for csum calculation when packet is read */
1035     elem->packet.payload = dst;
1036     dst += extraLen;
1037
1038     src = NdisGetDataBuffer(nb, dataLen, dst, 1, 0);
1039     if (src == NULL) {
1040         ovsUserStats.dropDuetoResource++;
1041         goto fail;
1042     }    else if (src != dst) {
1043         /* Copy the data from the NDIS buffer to dst. */
1044         RtlCopyMemory(dst, src, dataLen);
1045     }
1046
1047     /* Set csum if was offloaded */
1048     OvsCompletePacketHeader(dst, isRecv, csumInfo, hdrInfo, &elem->hdrInfo);
1049
1050     /*
1051      * Finally insert VLAN tag
1052      */
1053     if (extraLen) {
1054         dst = elem->packet.payload;
1055         src = dst + extraLen;
1056         ((UINT32 *)dst)[0] = ((UINT32 *)src)[0];
1057         ((UINT32 *)dst)[1] = ((UINT32 *)src)[1];
1058         ((UINT32 *)dst)[2] = ((UINT32 *)src)[2];
1059         dst += 12;
1060         ((UINT16 *)dst)[0] = htons(0x8100);
1061         ((UINT16 *)dst)[1] = htons(vlanInfo.TagHeader.VlanId |
1062             (vlanInfo.TagHeader.UserPriority << 13));
1063         elem->hdrInfo.l3Offset += VLAN_TAG_SIZE;
1064         elem->hdrInfo.l4Offset += VLAN_TAG_SIZE;
1065         ovsUserStats.vlanInsert++;
1066     }
1067     return elem;
1068 fail:
1069     OvsFreeMemory(elem);
1070     return NULL;
1071 }