drivers/infiniband/sw/rxe/rxe_resp.c

   1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
   2 /*
   3  * Copyright (c) 2016 Mellanox Technologies Ltd. All rights reserved.
   4  * Copyright (c) 2015 System Fabric Works, Inc. All rights reserved.
   5  */
   6
   7 #include <linux/skbuff.h>
   8
   9 #include "rxe.h"
  10 #include "rxe_loc.h"
  11 #include "rxe_queue.h"
  12
  13 enum resp_states {
  14         RESPST_NONE,
  15         RESPST_GET_REQ,
  16         RESPST_CHK_PSN,
  17         RESPST_CHK_OP_SEQ,
  18         RESPST_CHK_OP_VALID,
  19         RESPST_CHK_RESOURCE,
  20         RESPST_CHK_LENGTH,
  21         RESPST_CHK_RKEY,
  22         RESPST_EXECUTE,
  23         RESPST_READ_REPLY,
  24         RESPST_COMPLETE,
  25         RESPST_ACKNOWLEDGE,
  26         RESPST_CLEANUP,
  27         RESPST_DUPLICATE_REQUEST,
  28         RESPST_ERR_MALFORMED_WQE,
  29         RESPST_ERR_UNSUPPORTED_OPCODE,
  30         RESPST_ERR_MISALIGNED_ATOMIC,
  31         RESPST_ERR_PSN_OUT_OF_SEQ,
  32         RESPST_ERR_MISSING_OPCODE_FIRST,
  33         RESPST_ERR_MISSING_OPCODE_LAST_C,
  34         RESPST_ERR_MISSING_OPCODE_LAST_D1E,
  35         RESPST_ERR_TOO_MANY_RDMA_ATM_REQ,
  36         RESPST_ERR_RNR,
  37         RESPST_ERR_RKEY_VIOLATION,
  38         RESPST_ERR_INVALIDATE_RKEY,
  39         RESPST_ERR_LENGTH,
  40         RESPST_ERR_CQ_OVERFLOW,
  41         RESPST_ERROR,
  42         RESPST_RESET,
  43         RESPST_DONE,
  44         RESPST_EXIT,
  45 };
  46
  47 static char *resp_state_name[] = {
  48         [RESPST_NONE]                           = "NONE",
  49         [RESPST_GET_REQ]                        = "GET_REQ",
  50         [RESPST_CHK_PSN]                        = "CHK_PSN",
  51         [RESPST_CHK_OP_SEQ]                     = "CHK_OP_SEQ",
  52         [RESPST_CHK_OP_VALID]                   = "CHK_OP_VALID",
  53         [RESPST_CHK_RESOURCE]                   = "CHK_RESOURCE",
  54         [RESPST_CHK_LENGTH]                     = "CHK_LENGTH",
  55         [RESPST_CHK_RKEY]                       = "CHK_RKEY",
  56         [RESPST_EXECUTE]                        = "EXECUTE",
  57         [RESPST_READ_REPLY]                     = "READ_REPLY",
  58         [RESPST_COMPLETE]                       = "COMPLETE",
  59         [RESPST_ACKNOWLEDGE]                    = "ACKNOWLEDGE",
  60         [RESPST_CLEANUP]                        = "CLEANUP",
  61         [RESPST_DUPLICATE_REQUEST]              = "DUPLICATE_REQUEST",
  62         [RESPST_ERR_MALFORMED_WQE]              = "ERR_MALFORMED_WQE",
  63         [RESPST_ERR_UNSUPPORTED_OPCODE]         = "ERR_UNSUPPORTED_OPCODE",
  64         [RESPST_ERR_MISALIGNED_ATOMIC]          = "ERR_MISALIGNED_ATOMIC",
  65         [RESPST_ERR_PSN_OUT_OF_SEQ]             = "ERR_PSN_OUT_OF_SEQ",
  66         [RESPST_ERR_MISSING_OPCODE_FIRST]       = "ERR_MISSING_OPCODE_FIRST",
  67         [RESPST_ERR_MISSING_OPCODE_LAST_C]      = "ERR_MISSING_OPCODE_LAST_C",
  68         [RESPST_ERR_MISSING_OPCODE_LAST_D1E]    = "ERR_MISSING_OPCODE_LAST_D1E",
  69         [RESPST_ERR_TOO_MANY_RDMA_ATM_REQ]      = "ERR_TOO_MANY_RDMA_ATM_REQ",
  70         [RESPST_ERR_RNR]                        = "ERR_RNR",
  71         [RESPST_ERR_RKEY_VIOLATION]             = "ERR_RKEY_VIOLATION",
  72         [RESPST_ERR_INVALIDATE_RKEY]            = "ERR_INVALIDATE_RKEY_VIOLATION",
  73         [RESPST_ERR_LENGTH]                     = "ERR_LENGTH",
  74         [RESPST_ERR_CQ_OVERFLOW]                = "ERR_CQ_OVERFLOW",
  75         [RESPST_ERROR]                          = "ERROR",
  76         [RESPST_RESET]                          = "RESET",
  77         [RESPST_DONE]                           = "DONE",
  78         [RESPST_EXIT]                           = "EXIT",
  79 };
  80
  81 /* rxe_recv calls here to add a request packet to the input queue */
  82 void rxe_resp_queue_pkt(struct rxe_qp *qp, struct sk_buff *skb)
  83 {
  84         int must_sched;
  85         struct rxe_pkt_info *pkt = SKB_TO_PKT(skb);
  86
  87         skb_queue_tail(&qp->req_pkts, skb);
  88
  89         must_sched = (pkt->opcode == IB_OPCODE_RC_RDMA_READ_REQUEST) ||
  90                         (skb_queue_len(&qp->req_pkts) > 1);
  91
  92         rxe_run_task(&qp->resp.task, must_sched);
  93 }
  94
  95 static inline enum resp_states get_req(struct rxe_qp *qp,
  96                                        struct rxe_pkt_info **pkt_p)
  97 {
  98         struct sk_buff *skb;
  99
 100         if (qp->resp.state == QP_STATE_ERROR) {
 101                 while ((skb = skb_dequeue(&qp->req_pkts))) {
 102                         rxe_put(qp);
 103                         kfree_skb(skb);
 104                         ib_device_put(qp->ibqp.device);
 105                 }
 106
 107                 /* go drain recv wr queue */
 108                 return RESPST_CHK_RESOURCE;
 109         }
 110
 111         skb = skb_peek(&qp->req_pkts);
 112         if (!skb)
 113                 return RESPST_EXIT;
 114
 115         *pkt_p = SKB_TO_PKT(skb);
 116
 117         return (qp->resp.res) ? RESPST_READ_REPLY : RESPST_CHK_PSN;
 118 }
 119
 120 static enum resp_states check_psn(struct rxe_qp *qp,
 121                                   struct rxe_pkt_info *pkt)
 122 {
 123         int diff = psn_compare(pkt->psn, qp->resp.psn);
 124         struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
 125
 126         switch (qp_type(qp)) {
 127         case IB_QPT_RC:
 128                 if (diff > 0) {
 129                         if (qp->resp.sent_psn_nak)
 130                                 return RESPST_CLEANUP;
 131
 132                         qp->resp.sent_psn_nak = 1;
 133                         rxe_counter_inc(rxe, RXE_CNT_OUT_OF_SEQ_REQ);
 134                         return RESPST_ERR_PSN_OUT_OF_SEQ;
 135
 136                 } else if (diff < 0) {
 137                         rxe_counter_inc(rxe, RXE_CNT_DUP_REQ);
 138                         return RESPST_DUPLICATE_REQUEST;
 139                 }
 140
 141                 if (qp->resp.sent_psn_nak)
 142                         qp->resp.sent_psn_nak = 0;
 143
 144                 break;
 145
 146         case IB_QPT_UC:
 147                 if (qp->resp.drop_msg || diff != 0) {
 148                         if (pkt->mask & RXE_START_MASK) {
 149                                 qp->resp.drop_msg = 0;
 150                                 return RESPST_CHK_OP_SEQ;
 151                         }
 152
 153                         qp->resp.drop_msg = 1;
 154                         return RESPST_CLEANUP;
 155                 }
 156                 break;
 157         default:
 158                 break;
 159         }
 160
 161         return RESPST_CHK_OP_SEQ;
 162 }
 163
 164 static enum resp_states check_op_seq(struct rxe_qp *qp,
 165                                      struct rxe_pkt_info *pkt)
 166 {
 167         switch (qp_type(qp)) {
 168         case IB_QPT_RC:
 169                 switch (qp->resp.opcode) {
 170                 case IB_OPCODE_RC_SEND_FIRST:
 171                 case IB_OPCODE_RC_SEND_MIDDLE:
 172                         switch (pkt->opcode) {
 173                         case IB_OPCODE_RC_SEND_MIDDLE:
 174                         case IB_OPCODE_RC_SEND_LAST:
 175                         case IB_OPCODE_RC_SEND_LAST_WITH_IMMEDIATE:
 176                         case IB_OPCODE_RC_SEND_LAST_WITH_INVALIDATE:
 177                                 return RESPST_CHK_OP_VALID;
 178                         default:
 179                                 return RESPST_ERR_MISSING_OPCODE_LAST_C;
 180                         }
 181
 182                 case IB_OPCODE_RC_RDMA_WRITE_FIRST:
 183                 case IB_OPCODE_RC_RDMA_WRITE_MIDDLE:
 184                         switch (pkt->opcode) {
 185                         case IB_OPCODE_RC_RDMA_WRITE_MIDDLE:
 186                         case IB_OPCODE_RC_RDMA_WRITE_LAST:
 187                         case IB_OPCODE_RC_RDMA_WRITE_LAST_WITH_IMMEDIATE:
 188                                 return RESPST_CHK_OP_VALID;
 189                         default:
 190                                 return RESPST_ERR_MISSING_OPCODE_LAST_C;
 191                         }
 192
 193                 default:
 194                         switch (pkt->opcode) {
 195                         case IB_OPCODE_RC_SEND_MIDDLE:
 196                         case IB_OPCODE_RC_SEND_LAST:
 197                         case IB_OPCODE_RC_SEND_LAST_WITH_IMMEDIATE:
 198                         case IB_OPCODE_RC_SEND_LAST_WITH_INVALIDATE:
 199                         case IB_OPCODE_RC_RDMA_WRITE_MIDDLE:
 200                         case IB_OPCODE_RC_RDMA_WRITE_LAST:
 201                         case IB_OPCODE_RC_RDMA_WRITE_LAST_WITH_IMMEDIATE:
 202                                 return RESPST_ERR_MISSING_OPCODE_FIRST;
 203                         default:
 204                                 return RESPST_CHK_OP_VALID;
 205                         }
 206                 }
 207                 break;
 208
 209         case IB_QPT_UC:
 210                 switch (qp->resp.opcode) {
 211                 case IB_OPCODE_UC_SEND_FIRST:
 212                 case IB_OPCODE_UC_SEND_MIDDLE:
 213                         switch (pkt->opcode) {
 214                         case IB_OPCODE_UC_SEND_MIDDLE:
 215                         case IB_OPCODE_UC_SEND_LAST:
 216                         case IB_OPCODE_UC_SEND_LAST_WITH_IMMEDIATE:
 217                                 return RESPST_CHK_OP_VALID;
 218                         default:
 219                                 return RESPST_ERR_MISSING_OPCODE_LAST_D1E;
 220                         }
 221
 222                 case IB_OPCODE_UC_RDMA_WRITE_FIRST:
 223                 case IB_OPCODE_UC_RDMA_WRITE_MIDDLE:
 224                         switch (pkt->opcode) {
 225                         case IB_OPCODE_UC_RDMA_WRITE_MIDDLE:
 226                         case IB_OPCODE_UC_RDMA_WRITE_LAST:
 227                         case IB_OPCODE_UC_RDMA_WRITE_LAST_WITH_IMMEDIATE:
 228                                 return RESPST_CHK_OP_VALID;
 229                         default:
 230                                 return RESPST_ERR_MISSING_OPCODE_LAST_D1E;
 231                         }
 232
 233                 default:
 234                         switch (pkt->opcode) {
 235                         case IB_OPCODE_UC_SEND_MIDDLE:
 236                         case IB_OPCODE_UC_SEND_LAST:
 237                         case IB_OPCODE_UC_SEND_LAST_WITH_IMMEDIATE:
 238                         case IB_OPCODE_UC_RDMA_WRITE_MIDDLE:
 239                         case IB_OPCODE_UC_RDMA_WRITE_LAST:
 240                         case IB_OPCODE_UC_RDMA_WRITE_LAST_WITH_IMMEDIATE:
 241                                 qp->resp.drop_msg = 1;
 242                                 return RESPST_CLEANUP;
 243                         default:
 244                                 return RESPST_CHK_OP_VALID;
 245                         }
 246                 }
 247                 break;
 248
 249         default:
 250                 return RESPST_CHK_OP_VALID;
 251         }
 252 }
 253
 254 static enum resp_states check_op_valid(struct rxe_qp *qp,
 255                                        struct rxe_pkt_info *pkt)
 256 {
 257         switch (qp_type(qp)) {
 258         case IB_QPT_RC:
 259                 if (((pkt->mask & RXE_READ_MASK) &&
 260                      !(qp->attr.qp_access_flags & IB_ACCESS_REMOTE_READ)) ||
 261                     ((pkt->mask & RXE_WRITE_MASK) &&
 262                      !(qp->attr.qp_access_flags & IB_ACCESS_REMOTE_WRITE)) ||
 263                     ((pkt->mask & RXE_ATOMIC_MASK) &&
 264                      !(qp->attr.qp_access_flags & IB_ACCESS_REMOTE_ATOMIC))) {
 265                         return RESPST_ERR_UNSUPPORTED_OPCODE;
 266                 }
 267
 268                 break;
 269
 270         case IB_QPT_UC:
 271                 if ((pkt->mask & RXE_WRITE_MASK) &&
 272                     !(qp->attr.qp_access_flags & IB_ACCESS_REMOTE_WRITE)) {
 273                         qp->resp.drop_msg = 1;
 274                         return RESPST_CLEANUP;
 275                 }
 276
 277                 break;
 278
 279         case IB_QPT_UD:
 280         case IB_QPT_SMI:
 281         case IB_QPT_GSI:
 282                 break;
 283
 284         default:
 285                 WARN_ON_ONCE(1);
 286                 break;
 287         }
 288
 289         return RESPST_CHK_RESOURCE;
 290 }
 291
 292 static enum resp_states get_srq_wqe(struct rxe_qp *qp)
 293 {
 294         struct rxe_srq *srq = qp->srq;
 295         struct rxe_queue *q = srq->rq.queue;
 296         struct rxe_recv_wqe *wqe;
 297         struct ib_event ev;
 298         unsigned int count;
 299         size_t size;
 300         unsigned long flags;
 301
 302         if (srq->error)
 303                 return RESPST_ERR_RNR;
 304
 305         spin_lock_irqsave(&srq->rq.consumer_lock, flags);
 306
 307         wqe = queue_head(q, QUEUE_TYPE_FROM_CLIENT);
 308         if (!wqe) {
 309                 spin_unlock_irqrestore(&srq->rq.consumer_lock, flags);
 310                 return RESPST_ERR_RNR;
 311         }
 312
 313         /* don't trust user space data */
 314         if (unlikely(wqe->dma.num_sge > srq->rq.max_sge)) {
 315                 spin_unlock_irqrestore(&srq->rq.consumer_lock, flags);
 316                 pr_warn("%s: invalid num_sge in SRQ entry\n", __func__);
 317                 return RESPST_ERR_MALFORMED_WQE;
 318         }
 319         size = sizeof(*wqe) + wqe->dma.num_sge*sizeof(struct rxe_sge);
 320         memcpy(&qp->resp.srq_wqe, wqe, size);
 321
 322         qp->resp.wqe = &qp->resp.srq_wqe.wqe;
 323         queue_advance_consumer(q, QUEUE_TYPE_FROM_CLIENT);
 324         count = queue_count(q, QUEUE_TYPE_FROM_CLIENT);
 325
 326         if (srq->limit && srq->ibsrq.event_handler && (count < srq->limit)) {
 327                 srq->limit = 0;
 328                 goto event;
 329         }
 330
 331         spin_unlock_irqrestore(&srq->rq.consumer_lock, flags);
 332         return RESPST_CHK_LENGTH;
 333
 334 event:
 335         spin_unlock_irqrestore(&srq->rq.consumer_lock, flags);
 336         ev.device = qp->ibqp.device;
 337         ev.element.srq = qp->ibqp.srq;
 338         ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
 339         srq->ibsrq.event_handler(&ev, srq->ibsrq.srq_context);
 340         return RESPST_CHK_LENGTH;
 341 }
 342
 343 static enum resp_states check_resource(struct rxe_qp *qp,
 344                                        struct rxe_pkt_info *pkt)
 345 {
 346         struct rxe_srq *srq = qp->srq;
 347
 348         if (qp->resp.state == QP_STATE_ERROR) {
 349                 if (qp->resp.wqe) {
 350                         qp->resp.status = IB_WC_WR_FLUSH_ERR;
 351                         return RESPST_COMPLETE;
 352                 } else if (!srq) {
 353                         qp->resp.wqe = queue_head(qp->rq.queue,
 354                                         QUEUE_TYPE_FROM_CLIENT);
 355                         if (qp->resp.wqe) {
 356                                 qp->resp.status = IB_WC_WR_FLUSH_ERR;
 357                                 return RESPST_COMPLETE;
 358                         } else {
 359                                 return RESPST_EXIT;
 360                         }
 361                 } else {
 362                         return RESPST_EXIT;
 363                 }
 364         }
 365
 366         if (pkt->mask & RXE_READ_OR_ATOMIC_MASK) {
 367                 /* it is the requesters job to not send
 368                  * too many read/atomic ops, we just
 369                  * recycle the responder resource queue
 370                  */
 371                 if (likely(qp->attr.max_dest_rd_atomic > 0))
 372                         return RESPST_CHK_LENGTH;
 373                 else
 374                         return RESPST_ERR_TOO_MANY_RDMA_ATM_REQ;
 375         }
 376
 377         if (pkt->mask & RXE_RWR_MASK) {
 378                 if (srq)
 379                         return get_srq_wqe(qp);
 380
 381                 qp->resp.wqe = queue_head(qp->rq.queue,
 382                                 QUEUE_TYPE_FROM_CLIENT);
 383                 return (qp->resp.wqe) ? RESPST_CHK_LENGTH : RESPST_ERR_RNR;
 384         }
 385
 386         return RESPST_CHK_LENGTH;
 387 }
 388
 389 static enum resp_states check_length(struct rxe_qp *qp,
 390                                      struct rxe_pkt_info *pkt)
 391 {
 392         switch (qp_type(qp)) {
 393         case IB_QPT_RC:
 394                 return RESPST_CHK_RKEY;
 395
 396         case IB_QPT_UC:
 397                 return RESPST_CHK_RKEY;
 398
 399         default:
 400                 return RESPST_CHK_RKEY;
 401         }
 402 }
 403
 404 static enum resp_states check_rkey(struct rxe_qp *qp,
 405                                    struct rxe_pkt_info *pkt)
 406 {
 407         struct rxe_mr *mr = NULL;
 408         struct rxe_mw *mw = NULL;
 409         u64 va;
 410         u32 rkey;
 411         u32 resid;
 412         u32 pktlen;
 413         int mtu = qp->mtu;
 414         enum resp_states state;
 415         int access;
 416
 417         if (pkt->mask & RXE_READ_OR_WRITE_MASK) {
 418                 if (pkt->mask & RXE_RETH_MASK) {
 419                         qp->resp.va = reth_va(pkt);
 420                         qp->resp.offset = 0;
 421                         qp->resp.rkey = reth_rkey(pkt);
 422                         qp->resp.resid = reth_len(pkt);
 423                         qp->resp.length = reth_len(pkt);
 424                 }
 425                 access = (pkt->mask & RXE_READ_MASK) ? IB_ACCESS_REMOTE_READ
 426                                                      : IB_ACCESS_REMOTE_WRITE;
 427         } else if (pkt->mask & RXE_ATOMIC_MASK) {
 428                 qp->resp.va = atmeth_va(pkt);
 429                 qp->resp.offset = 0;
 430                 qp->resp.rkey = atmeth_rkey(pkt);
 431                 qp->resp.resid = sizeof(u64);
 432                 access = IB_ACCESS_REMOTE_ATOMIC;
 433         } else {
 434                 return RESPST_EXECUTE;
 435         }
 436
 437         /* A zero-byte op is not required to set an addr or rkey. */
 438         if ((pkt->mask & RXE_READ_OR_WRITE_MASK) &&
 439             (pkt->mask & RXE_RETH_MASK) &&
 440             reth_len(pkt) == 0) {
 441                 return RESPST_EXECUTE;
 442         }
 443
 444         va      = qp->resp.va;
 445         rkey    = qp->resp.rkey;
 446         resid   = qp->resp.resid;
 447         pktlen  = payload_size(pkt);
 448
 449         if (rkey_is_mw(rkey)) {
 450                 mw = rxe_lookup_mw(qp, access, rkey);
 451                 if (!mw) {
 452                         pr_err("%s: no MW matches rkey %#x\n", __func__, rkey);
 453                         state = RESPST_ERR_RKEY_VIOLATION;
 454                         goto err;
 455                 }
 456
 457                 mr = mw->mr;
 458                 if (!mr) {
 459                         pr_err("%s: MW doesn't have an MR\n", __func__);
 460                         state = RESPST_ERR_RKEY_VIOLATION;
 461                         goto err;
 462                 }
 463
 464                 if (mw->access & IB_ZERO_BASED)
 465                         qp->resp.offset = mw->addr;
 466
 467                 rxe_put(mw);
 468                 rxe_get(mr);
 469         } else {
 470                 mr = lookup_mr(qp->pd, access, rkey, RXE_LOOKUP_REMOTE);
 471                 if (!mr) {
 472                         pr_err("%s: no MR matches rkey %#x\n", __func__, rkey);
 473                         state = RESPST_ERR_RKEY_VIOLATION;
 474                         goto err;
 475                 }
 476         }
 477
 478         if (mr_check_range(mr, va + qp->resp.offset, resid)) {
 479                 state = RESPST_ERR_RKEY_VIOLATION;
 480                 goto err;
 481         }
 482
 483         if (pkt->mask & RXE_WRITE_MASK)  {
 484                 if (resid > mtu) {
 485                         if (pktlen != mtu || bth_pad(pkt)) {
 486                                 state = RESPST_ERR_LENGTH;
 487                                 goto err;
 488                         }
 489                 } else {
 490                         if (pktlen != resid) {
 491                                 state = RESPST_ERR_LENGTH;
 492                                 goto err;
 493                         }
 494                         if ((bth_pad(pkt) != (0x3 & (-resid)))) {
 495                                 /* This case may not be exactly that
 496                                  * but nothing else fits.
 497                                  */
 498                                 state = RESPST_ERR_LENGTH;
 499                                 goto err;
 500                         }
 501                 }
 502         }
 503
 504         WARN_ON_ONCE(qp->resp.mr);
 505
 506         qp->resp.mr = mr;
 507         return RESPST_EXECUTE;
 508
 509 err:
 510         if (mr)
 511                 rxe_put(mr);
 512         if (mw)
 513                 rxe_put(mw);
 514
 515         return state;
 516 }
 517
 518 static enum resp_states send_data_in(struct rxe_qp *qp, void *data_addr,
 519                                      int data_len)
 520 {
 521         int err;
 522
 523         err = copy_data(qp->pd, IB_ACCESS_LOCAL_WRITE, &qp->resp.wqe->dma,
 524                         data_addr, data_len, RXE_TO_MR_OBJ);
 525         if (unlikely(err))
 526                 return (err == -ENOSPC) ? RESPST_ERR_LENGTH
 527                                         : RESPST_ERR_MALFORMED_WQE;
 528
 529         return RESPST_NONE;
 530 }
 531
 532 static enum resp_states write_data_in(struct rxe_qp *qp,
 533                                       struct rxe_pkt_info *pkt)
 534 {
 535         enum resp_states rc = RESPST_NONE;
 536         int     err;
 537         int data_len = payload_size(pkt);
 538
 539         err = rxe_mr_copy(qp->resp.mr, qp->resp.va + qp->resp.offset,
 540                           payload_addr(pkt), data_len, RXE_TO_MR_OBJ);
 541         if (err) {
 542                 rc = RESPST_ERR_RKEY_VIOLATION;
 543                 goto out;
 544         }
 545
 546         qp->resp.va += data_len;
 547         qp->resp.resid -= data_len;
 548
 549 out:
 550         return rc;
 551 }
 552
 553 /* Guarantee atomicity of atomic operations at the machine level. */
 554 static DEFINE_SPINLOCK(atomic_ops_lock);
 555
 556 static enum resp_states process_atomic(struct rxe_qp *qp,
 557                                        struct rxe_pkt_info *pkt)
 558 {
 559         u64 *vaddr;
 560         enum resp_states ret;
 561         struct rxe_mr *mr = qp->resp.mr;
 562
 563         if (mr->state != RXE_MR_STATE_VALID) {
 564                 ret = RESPST_ERR_RKEY_VIOLATION;
 565                 goto out;
 566         }
 567
 568         vaddr = iova_to_vaddr(mr, qp->resp.va + qp->resp.offset, sizeof(u64));
 569
 570         /* check vaddr is 8 bytes aligned. */
 571         if (!vaddr || (uintptr_t)vaddr & 7) {
 572                 ret = RESPST_ERR_MISALIGNED_ATOMIC;
 573                 goto out;
 574         }
 575
 576         spin_lock_bh(&atomic_ops_lock);
 577
 578         qp->resp.atomic_orig = *vaddr;
 579
 580         if (pkt->opcode == IB_OPCODE_RC_COMPARE_SWAP ||
 581             pkt->opcode == IB_OPCODE_RD_COMPARE_SWAP) {
 582                 if (*vaddr == atmeth_comp(pkt))
 583                         *vaddr = atmeth_swap_add(pkt);
 584         } else {
 585                 *vaddr += atmeth_swap_add(pkt);
 586         }
 587
 588         spin_unlock_bh(&atomic_ops_lock);
 589
 590         ret = RESPST_NONE;
 591 out:
 592         return ret;
 593 }
 594
 595 static struct sk_buff *prepare_ack_packet(struct rxe_qp *qp,
 596                                           struct rxe_pkt_info *pkt,
 597                                           struct rxe_pkt_info *ack,
 598                                           int opcode,
 599                                           int payload,
 600                                           u32 psn,
 601                                           u8 syndrome)
 602 {
 603         struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
 604         struct sk_buff *skb;
 605         int paylen;
 606         int pad;
 607         int err;
 608
 609         /*
 610          * allocate packet
 611          */
 612         pad = (-payload) & 0x3;
 613         paylen = rxe_opcode[opcode].length + payload + pad + RXE_ICRC_SIZE;
 614
 615         skb = rxe_init_packet(rxe, &qp->pri_av, paylen, ack);
 616         if (!skb)
 617                 return NULL;
 618
 619         ack->qp = qp;
 620         ack->opcode = opcode;
 621         ack->mask = rxe_opcode[opcode].mask;
 622         ack->paylen = paylen;
 623         ack->psn = psn;
 624
 625         bth_init(ack, opcode, 0, 0, pad, IB_DEFAULT_PKEY_FULL,
 626                  qp->attr.dest_qp_num, 0, psn);
 627
 628         if (ack->mask & RXE_AETH_MASK) {
 629                 aeth_set_syn(ack, syndrome);
 630                 aeth_set_msn(ack, qp->resp.msn);
 631         }
 632
 633         if (ack->mask & RXE_ATMACK_MASK)
 634                 atmack_set_orig(ack, qp->resp.atomic_orig);
 635
 636         err = rxe_prepare(&qp->pri_av, ack, skb);
 637         if (err) {
 638                 kfree_skb(skb);
 639                 return NULL;
 640         }
 641
 642         return skb;
 643 }
 644
 645 static struct resp_res *rxe_prepare_read_res(struct rxe_qp *qp,
 646                                         struct rxe_pkt_info *pkt)
 647 {
 648         struct resp_res *res;
 649         u32 pkts;
 650
 651         res = &qp->resp.resources[qp->resp.res_head];
 652         rxe_advance_resp_resource(qp);
 653         free_rd_atomic_resource(qp, res);
 654
 655         res->type = RXE_READ_MASK;
 656         res->replay = 0;
 657         res->read.va = qp->resp.va + qp->resp.offset;
 658         res->read.va_org = qp->resp.va + qp->resp.offset;
 659         res->read.resid = qp->resp.resid;
 660         res->read.length = qp->resp.resid;
 661         res->read.rkey = qp->resp.rkey;
 662
 663         pkts = max_t(u32, (reth_len(pkt) + qp->mtu - 1)/qp->mtu, 1);
 664         res->first_psn = pkt->psn;
 665         res->cur_psn = pkt->psn;
 666         res->last_psn = (pkt->psn + pkts - 1) & BTH_PSN_MASK;
 667
 668         res->state = rdatm_res_state_new;
 669
 670         return res;
 671 }
 672
 673 /**
 674  * rxe_recheck_mr - revalidate MR from rkey and get a reference
 675  * @qp: the qp
 676  * @rkey: the rkey
 677  *
 678  * This code allows the MR to be invalidated or deregistered or
 679  * the MW if one was used to be invalidated or deallocated.
 680  * It is assumed that the access permissions if originally good
 681  * are OK and the mappings to be unchanged.
 682  *
 683  * TODO: If someone reregisters an MR to change its size or
 684  * access permissions during the processing of an RDMA read
 685  * we should kill the responder resource and complete the
 686  * operation with an error.
 687  *
 688  * Return: mr on success else NULL
 689  */
 690 static struct rxe_mr *rxe_recheck_mr(struct rxe_qp *qp, u32 rkey)
 691 {
 692         struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
 693         struct rxe_mr *mr;
 694         struct rxe_mw *mw;
 695
 696         if (rkey_is_mw(rkey)) {
 697                 mw = rxe_pool_get_index(&rxe->mw_pool, rkey >> 8);
 698                 if (!mw)
 699                         return NULL;
 700
 701                 mr = mw->mr;
 702                 if (mw->rkey != rkey || mw->state != RXE_MW_STATE_VALID ||
 703                     !mr || mr->state != RXE_MR_STATE_VALID) {
 704                         rxe_put(mw);
 705                         return NULL;
 706                 }
 707
 708                 rxe_get(mr);
 709                 rxe_put(mw);
 710
 711                 return mr;
 712         }
 713
 714         mr = rxe_pool_get_index(&rxe->mr_pool, rkey >> 8);
 715         if (!mr)
 716                 return NULL;
 717
 718         if (mr->rkey != rkey || mr->state != RXE_MR_STATE_VALID) {
 719                 rxe_put(mr);
 720                 return NULL;
 721         }
 722
 723         return mr;
 724 }
 725
 726 /* RDMA read response. If res is not NULL, then we have a current RDMA request
 727  * being processed or replayed.
 728  */
 729 static enum resp_states read_reply(struct rxe_qp *qp,
 730                                    struct rxe_pkt_info *req_pkt)
 731 {
 732         struct rxe_pkt_info ack_pkt;
 733         struct sk_buff *skb;
 734         int mtu = qp->mtu;
 735         enum resp_states state;
 736         int payload;
 737         int opcode;
 738         int err;
 739         struct resp_res *res = qp->resp.res;
 740         struct rxe_mr *mr;
 741
 742         if (!res) {
 743                 res = rxe_prepare_read_res(qp, req_pkt);
 744                 qp->resp.res = res;
 745         }
 746
 747         if (res->state == rdatm_res_state_new) {
 748                 if (!res->replay) {
 749                         mr = qp->resp.mr;
 750                         qp->resp.mr = NULL;
 751                 } else {
 752                         mr = rxe_recheck_mr(qp, res->read.rkey);
 753                         if (!mr)
 754                                 return RESPST_ERR_RKEY_VIOLATION;
 755                 }
 756
 757                 if (res->read.resid <= mtu)
 758                         opcode = IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY;
 759                 else
 760                         opcode = IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST;
 761         } else {
 762                 mr = rxe_recheck_mr(qp, res->read.rkey);
 763                 if (!mr)
 764                         return RESPST_ERR_RKEY_VIOLATION;
 765
 766                 if (res->read.resid > mtu)
 767                         opcode = IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE;
 768                 else
 769                         opcode = IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST;
 770         }
 771
 772         res->state = rdatm_res_state_next;
 773
 774         payload = min_t(int, res->read.resid, mtu);
 775
 776         skb = prepare_ack_packet(qp, req_pkt, &ack_pkt, opcode, payload,
 777                                  res->cur_psn, AETH_ACK_UNLIMITED);
 778         if (!skb)
 779                 return RESPST_ERR_RNR;
 780
 781         err = rxe_mr_copy(mr, res->read.va, payload_addr(&ack_pkt),
 782                           payload, RXE_FROM_MR_OBJ);
 783         if (err)
 784                 pr_err("Failed copying memory\n");
 785         if (mr)
 786                 rxe_put(mr);
 787
 788         if (bth_pad(&ack_pkt)) {
 789                 u8 *pad = payload_addr(&ack_pkt) + payload;
 790
 791                 memset(pad, 0, bth_pad(&ack_pkt));
 792         }
 793
 794         err = rxe_xmit_packet(qp, &ack_pkt, skb);
 795         if (err) {
 796                 pr_err("Failed sending RDMA reply.\n");
 797                 return RESPST_ERR_RNR;
 798         }
 799
 800         res->read.va += payload;
 801         res->read.resid -= payload;
 802         res->cur_psn = (res->cur_psn + 1) & BTH_PSN_MASK;
 803
 804         if (res->read.resid > 0) {
 805                 state = RESPST_DONE;
 806         } else {
 807                 qp->resp.res = NULL;
 808                 if (!res->replay)
 809                         qp->resp.opcode = -1;
 810                 if (psn_compare(res->cur_psn, qp->resp.psn) >= 0)
 811                         qp->resp.psn = res->cur_psn;
 812                 state = RESPST_CLEANUP;
 813         }
 814
 815         return state;
 816 }
 817
 818 static int invalidate_rkey(struct rxe_qp *qp, u32 rkey)
 819 {
 820         if (rkey_is_mw(rkey))
 821                 return rxe_invalidate_mw(qp, rkey);
 822         else
 823                 return rxe_invalidate_mr(qp, rkey);
 824 }
 825
 826 /* Executes a new request. A retried request never reach that function (send
 827  * and writes are discarded, and reads and atomics are retried elsewhere.
 828  */
 829 static enum resp_states execute(struct rxe_qp *qp, struct rxe_pkt_info *pkt)
 830 {
 831         enum resp_states err;
 832         struct sk_buff *skb = PKT_TO_SKB(pkt);
 833         union rdma_network_hdr hdr;
 834
 835         if (pkt->mask & RXE_SEND_MASK) {
 836                 if (qp_type(qp) == IB_QPT_UD ||
 837                     qp_type(qp) == IB_QPT_SMI ||
 838                     qp_type(qp) == IB_QPT_GSI) {
 839                         if (skb->protocol == htons(ETH_P_IP)) {
 840                                 memset(&hdr.reserved, 0,
 841                                                 sizeof(hdr.reserved));
 842                                 memcpy(&hdr.roce4grh, ip_hdr(skb),
 843                                                 sizeof(hdr.roce4grh));
 844                                 err = send_data_in(qp, &hdr, sizeof(hdr));
 845                         } else {
 846                                 err = send_data_in(qp, ipv6_hdr(skb),
 847                                                 sizeof(hdr));
 848                         }
 849                         if (err)
 850                                 return err;
 851                 }
 852                 err = send_data_in(qp, payload_addr(pkt), payload_size(pkt));
 853                 if (err)
 854                         return err;
 855         } else if (pkt->mask & RXE_WRITE_MASK) {
 856                 err = write_data_in(qp, pkt);
 857                 if (err)
 858                         return err;
 859         } else if (pkt->mask & RXE_READ_MASK) {
 860                 /* For RDMA Read we can increment the msn now. See C9-148. */
 861                 qp->resp.msn++;
 862                 return RESPST_READ_REPLY;
 863         } else if (pkt->mask & RXE_ATOMIC_MASK) {
 864                 err = process_atomic(qp, pkt);
 865                 if (err)
 866                         return err;
 867         } else {
 868                 /* Unreachable */
 869                 WARN_ON_ONCE(1);
 870         }
 871
 872         if (pkt->mask & RXE_IETH_MASK) {
 873                 u32 rkey = ieth_rkey(pkt);
 874
 875                 err = invalidate_rkey(qp, rkey);
 876                 if (err)
 877                         return RESPST_ERR_INVALIDATE_RKEY;
 878         }
 879
 880         if (pkt->mask & RXE_END_MASK)
 881                 /* We successfully processed this new request. */
 882                 qp->resp.msn++;
 883
 884         /* next expected psn, read handles this separately */
 885         qp->resp.psn = (pkt->psn + 1) & BTH_PSN_MASK;
 886         qp->resp.ack_psn = qp->resp.psn;
 887
 888         qp->resp.opcode = pkt->opcode;
 889         qp->resp.status = IB_WC_SUCCESS;
 890
 891         if (pkt->mask & RXE_COMP_MASK)
 892                 return RESPST_COMPLETE;
 893         else if (qp_type(qp) == IB_QPT_RC)
 894                 return RESPST_ACKNOWLEDGE;
 895         else
 896                 return RESPST_CLEANUP;
 897 }
 898
 899 static enum resp_states do_complete(struct rxe_qp *qp,
 900                                     struct rxe_pkt_info *pkt)
 901 {
 902         struct rxe_cqe cqe;
 903         struct ib_wc *wc = &cqe.ibwc;
 904         struct ib_uverbs_wc *uwc = &cqe.uibwc;
 905         struct rxe_recv_wqe *wqe = qp->resp.wqe;
 906         struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
 907
 908         if (!wqe)
 909                 goto finish;
 910
 911         memset(&cqe, 0, sizeof(cqe));
 912
 913         if (qp->rcq->is_user) {
 914                 uwc->status             = qp->resp.status;
 915                 uwc->qp_num             = qp->ibqp.qp_num;
 916                 uwc->wr_id              = wqe->wr_id;
 917         } else {
 918                 wc->status              = qp->resp.status;
 919                 wc->qp                  = &qp->ibqp;
 920                 wc->wr_id               = wqe->wr_id;
 921         }
 922
 923         if (wc->status == IB_WC_SUCCESS) {
 924                 rxe_counter_inc(rxe, RXE_CNT_RDMA_RECV);
 925                 wc->opcode = (pkt->mask & RXE_IMMDT_MASK &&
 926                                 pkt->mask & RXE_WRITE_MASK) ?
 927                                         IB_WC_RECV_RDMA_WITH_IMM : IB_WC_RECV;
 928                 wc->byte_len = (pkt->mask & RXE_IMMDT_MASK &&
 929                                 pkt->mask & RXE_WRITE_MASK) ?
 930                                         qp->resp.length : wqe->dma.length - wqe->dma.resid;
 931
 932                 /* fields after byte_len are different between kernel and user
 933                  * space
 934                  */
 935                 if (qp->rcq->is_user) {
 936                         uwc->wc_flags = IB_WC_GRH;
 937
 938                         if (pkt->mask & RXE_IMMDT_MASK) {
 939                                 uwc->wc_flags |= IB_WC_WITH_IMM;
 940                                 uwc->ex.imm_data = immdt_imm(pkt);
 941                         }
 942
 943                         if (pkt->mask & RXE_IETH_MASK) {
 944                                 uwc->wc_flags |= IB_WC_WITH_INVALIDATE;
 945                                 uwc->ex.invalidate_rkey = ieth_rkey(pkt);
 946                         }
 947
 948                         if (pkt->mask & RXE_DETH_MASK)
 949                                 uwc->src_qp = deth_sqp(pkt);
 950
 951                         uwc->port_num           = qp->attr.port_num;
 952                 } else {
 953                         struct sk_buff *skb = PKT_TO_SKB(pkt);
 954
 955                         wc->wc_flags = IB_WC_GRH | IB_WC_WITH_NETWORK_HDR_TYPE;
 956                         if (skb->protocol == htons(ETH_P_IP))
 957                                 wc->network_hdr_type = RDMA_NETWORK_IPV4;
 958                         else
 959                                 wc->network_hdr_type = RDMA_NETWORK_IPV6;
 960
 961                         if (is_vlan_dev(skb->dev)) {
 962                                 wc->wc_flags |= IB_WC_WITH_VLAN;
 963                                 wc->vlan_id = vlan_dev_vlan_id(skb->dev);
 964                         }
 965
 966                         if (pkt->mask & RXE_IMMDT_MASK) {
 967                                 wc->wc_flags |= IB_WC_WITH_IMM;
 968                                 wc->ex.imm_data = immdt_imm(pkt);
 969                         }
 970
 971                         if (pkt->mask & RXE_IETH_MASK) {
 972                                 wc->wc_flags |= IB_WC_WITH_INVALIDATE;
 973                                 wc->ex.invalidate_rkey = ieth_rkey(pkt);
 974                         }
 975
 976                         if (pkt->mask & RXE_DETH_MASK)
 977                                 wc->src_qp = deth_sqp(pkt);
 978
 979                         wc->port_num            = qp->attr.port_num;
 980                 }
 981         }
 982
 983         /* have copy for srq and reference for !srq */
 984         if (!qp->srq)
 985                 queue_advance_consumer(qp->rq.queue, QUEUE_TYPE_FROM_CLIENT);
 986
 987         qp->resp.wqe = NULL;
 988
 989         if (rxe_cq_post(qp->rcq, &cqe, pkt ? bth_se(pkt) : 1))
 990                 return RESPST_ERR_CQ_OVERFLOW;
 991
 992 finish:
 993         if (unlikely(qp->resp.state == QP_STATE_ERROR))
 994                 return RESPST_CHK_RESOURCE;
 995         if (unlikely(!pkt))
 996                 return RESPST_DONE;
 997         if (qp_type(qp) == IB_QPT_RC)
 998                 return RESPST_ACKNOWLEDGE;
 999         else
1000                 return RESPST_CLEANUP;
1001 }
1002
1003 static int send_ack(struct rxe_qp *qp, struct rxe_pkt_info *pkt,
1004                     u8 syndrome, u32 psn)
1005 {
1006         int err = 0;
1007         struct rxe_pkt_info ack_pkt;
1008         struct sk_buff *skb;
1009
1010         skb = prepare_ack_packet(qp, pkt, &ack_pkt, IB_OPCODE_RC_ACKNOWLEDGE,
1011                                  0, psn, syndrome);
1012         if (!skb) {
1013                 err = -ENOMEM;
1014                 goto err1;
1015         }
1016
1017         err = rxe_xmit_packet(qp, &ack_pkt, skb);
1018         if (err)
1019                 pr_err_ratelimited("Failed sending ack\n");
1020
1021 err1:
1022         return err;
1023 }
1024
1025 static int send_atomic_ack(struct rxe_qp *qp, struct rxe_pkt_info *pkt,
1026                            u8 syndrome)
1027 {
1028         int rc = 0;
1029         struct rxe_pkt_info ack_pkt;
1030         struct sk_buff *skb;
1031         struct resp_res *res;
1032
1033         skb = prepare_ack_packet(qp, pkt, &ack_pkt,
1034                                  IB_OPCODE_RC_ATOMIC_ACKNOWLEDGE, 0, pkt->psn,
1035                                  syndrome);
1036         if (!skb) {
1037                 rc = -ENOMEM;
1038                 goto out;
1039         }
1040
1041         res = &qp->resp.resources[qp->resp.res_head];
1042         free_rd_atomic_resource(qp, res);
1043         rxe_advance_resp_resource(qp);
1044
1045         skb_get(skb);
1046         res->type = RXE_ATOMIC_MASK;
1047         res->atomic.skb = skb;
1048         res->first_psn = ack_pkt.psn;
1049         res->last_psn  = ack_pkt.psn;
1050         res->cur_psn   = ack_pkt.psn;
1051
1052         rc = rxe_xmit_packet(qp, &ack_pkt, skb);
1053         if (rc) {
1054                 pr_err_ratelimited("Failed sending ack\n");
1055                 rxe_put(qp);
1056         }
1057 out:
1058         return rc;
1059 }
1060
1061 static enum resp_states acknowledge(struct rxe_qp *qp,
1062                                     struct rxe_pkt_info *pkt)
1063 {
1064         if (qp_type(qp) != IB_QPT_RC)
1065                 return RESPST_CLEANUP;
1066
1067         if (qp->resp.aeth_syndrome != AETH_ACK_UNLIMITED)
1068                 send_ack(qp, pkt, qp->resp.aeth_syndrome, pkt->psn);
1069         else if (pkt->mask & RXE_ATOMIC_MASK)
1070                 send_atomic_ack(qp, pkt, AETH_ACK_UNLIMITED);
1071         else if (bth_ack(pkt))
1072                 send_ack(qp, pkt, AETH_ACK_UNLIMITED, pkt->psn);
1073
1074         return RESPST_CLEANUP;
1075 }
1076
1077 static enum resp_states cleanup(struct rxe_qp *qp,
1078                                 struct rxe_pkt_info *pkt)
1079 {
1080         struct sk_buff *skb;
1081
1082         if (pkt) {
1083                 skb = skb_dequeue(&qp->req_pkts);
1084                 rxe_put(qp);
1085                 kfree_skb(skb);
1086                 ib_device_put(qp->ibqp.device);
1087         }
1088
1089         if (qp->resp.mr) {
1090                 rxe_put(qp->resp.mr);
1091                 qp->resp.mr = NULL;
1092         }
1093
1094         return RESPST_DONE;
1095 }
1096
1097 static struct resp_res *find_resource(struct rxe_qp *qp, u32 psn)
1098 {
1099         int i;
1100
1101         for (i = 0; i < qp->attr.max_dest_rd_atomic; i++) {
1102                 struct resp_res *res = &qp->resp.resources[i];
1103
1104                 if (res->type == 0)
1105                         continue;
1106
1107                 if (psn_compare(psn, res->first_psn) >= 0 &&
1108                     psn_compare(psn, res->last_psn) <= 0) {
1109                         return res;
1110                 }
1111         }
1112
1113         return NULL;
1114 }
1115
1116 static enum resp_states duplicate_request(struct rxe_qp *qp,
1117                                           struct rxe_pkt_info *pkt)
1118 {
1119         enum resp_states rc;
1120         u32 prev_psn = (qp->resp.ack_psn - 1) & BTH_PSN_MASK;
1121
1122         if (pkt->mask & RXE_SEND_MASK ||
1123             pkt->mask & RXE_WRITE_MASK) {
1124                 /* SEND. Ack again and cleanup. C9-105. */
1125                 send_ack(qp, pkt, AETH_ACK_UNLIMITED, prev_psn);
1126                 return RESPST_CLEANUP;
1127         } else if (pkt->mask & RXE_READ_MASK) {
1128                 struct resp_res *res;
1129
1130                 res = find_resource(qp, pkt->psn);
1131                 if (!res) {
1132                         /* Resource not found. Class D error.  Drop the
1133                          * request.
1134                          */
1135                         rc = RESPST_CLEANUP;
1136                         goto out;
1137                 } else {
1138                         /* Ensure this new request is the same as the previous
1139                          * one or a subset of it.
1140                          */
1141                         u64 iova = reth_va(pkt);
1142                         u32 resid = reth_len(pkt);
1143
1144                         if (iova < res->read.va_org ||
1145                             resid > res->read.length ||
1146                             (iova + resid) > (res->read.va_org +
1147                                               res->read.length)) {
1148                                 rc = RESPST_CLEANUP;
1149                                 goto out;
1150                         }
1151
1152                         if (reth_rkey(pkt) != res->read.rkey) {
1153                                 rc = RESPST_CLEANUP;
1154                                 goto out;
1155                         }
1156
1157                         res->cur_psn = pkt->psn;
1158                         res->state = (pkt->psn == res->first_psn) ?
1159                                         rdatm_res_state_new :
1160                                         rdatm_res_state_replay;
1161                         res->replay = 1;
1162
1163                         /* Reset the resource, except length. */
1164                         res->read.va_org = iova;
1165                         res->read.va = iova;
1166                         res->read.resid = resid;
1167
1168                         /* Replay the RDMA read reply. */
1169                         qp->resp.res = res;
1170                         rc = RESPST_READ_REPLY;
1171                         goto out;
1172                 }
1173         } else {
1174                 struct resp_res *res;
1175
1176                 /* Find the operation in our list of responder resources. */
1177                 res = find_resource(qp, pkt->psn);
1178                 if (res) {
1179                         skb_get(res->atomic.skb);
1180                         /* Resend the result. */
1181                         rc = rxe_xmit_packet(qp, pkt, res->atomic.skb);
1182                         if (rc) {
1183                                 pr_err("Failed resending result. This flow is not handled - skb ignored\n");
1184                                 rc = RESPST_CLEANUP;
1185                                 goto out;
1186                         }
1187                 }
1188
1189                 /* Resource not found. Class D error. Drop the request. */
1190                 rc = RESPST_CLEANUP;
1191                 goto out;
1192         }
1193 out:
1194         return rc;
1195 }
1196
1197 /* Process a class A or C. Both are treated the same in this implementation. */
1198 static void do_class_ac_error(struct rxe_qp *qp, u8 syndrome,
1199                               enum ib_wc_status status)
1200 {
1201         qp->resp.aeth_syndrome  = syndrome;
1202         qp->resp.status         = status;
1203
1204         /* indicate that we should go through the ERROR state */
1205         qp->resp.goto_error     = 1;
1206 }
1207
1208 static enum resp_states do_class_d1e_error(struct rxe_qp *qp)
1209 {
1210         /* UC */
1211         if (qp->srq) {
1212                 /* Class E */
1213                 qp->resp.drop_msg = 1;
1214                 if (qp->resp.wqe) {
1215                         qp->resp.status = IB_WC_REM_INV_REQ_ERR;
1216                         return RESPST_COMPLETE;
1217                 } else {
1218                         return RESPST_CLEANUP;
1219                 }
1220         } else {
1221                 /* Class D1. This packet may be the start of a
1222                  * new message and could be valid. The previous
1223                  * message is invalid and ignored. reset the
1224                  * recv wr to its original state
1225                  */
1226                 if (qp->resp.wqe) {
1227                         qp->resp.wqe->dma.resid = qp->resp.wqe->dma.length;
1228                         qp->resp.wqe->dma.cur_sge = 0;
1229                         qp->resp.wqe->dma.sge_offset = 0;
1230                         qp->resp.opcode = -1;
1231                 }
1232
1233                 if (qp->resp.mr) {
1234                         rxe_put(qp->resp.mr);
1235                         qp->resp.mr = NULL;
1236                 }
1237
1238                 return RESPST_CLEANUP;
1239         }
1240 }
1241
1242 static void rxe_drain_req_pkts(struct rxe_qp *qp, bool notify)
1243 {
1244         struct sk_buff *skb;
1245         struct rxe_queue *q = qp->rq.queue;
1246
1247         while ((skb = skb_dequeue(&qp->req_pkts))) {
1248                 rxe_put(qp);
1249                 kfree_skb(skb);
1250                 ib_device_put(qp->ibqp.device);
1251         }
1252
1253         if (notify)
1254                 return;
1255
1256         while (!qp->srq && q && queue_head(q, q->type))
1257                 queue_advance_consumer(q, q->type);
1258 }
1259
1260 int rxe_responder(void *arg)
1261 {
1262         struct rxe_qp *qp = (struct rxe_qp *)arg;
1263         struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
1264         enum resp_states state;
1265         struct rxe_pkt_info *pkt = NULL;
1266         int ret = 0;
1267
1268         rxe_get(qp);
1269
1270         qp->resp.aeth_syndrome = AETH_ACK_UNLIMITED;
1271
1272         if (!qp->valid) {
1273                 ret = -EINVAL;
1274                 goto done;
1275         }
1276
1277         switch (qp->resp.state) {
1278         case QP_STATE_RESET:
1279                 state = RESPST_RESET;
1280                 break;
1281
1282         default:
1283                 state = RESPST_GET_REQ;
1284                 break;
1285         }
1286
1287         while (1) {
1288                 pr_debug("qp#%d state = %s\n", qp_num(qp),
1289                          resp_state_name[state]);
1290                 switch (state) {
1291                 case RESPST_GET_REQ:
1292                         state = get_req(qp, &pkt);
1293                         break;
1294                 case RESPST_CHK_PSN:
1295                         state = check_psn(qp, pkt);
1296                         break;
1297                 case RESPST_CHK_OP_SEQ:
1298                         state = check_op_seq(qp, pkt);
1299                         break;
1300                 case RESPST_CHK_OP_VALID:
1301                         state = check_op_valid(qp, pkt);
1302                         break;
1303                 case RESPST_CHK_RESOURCE:
1304                         state = check_resource(qp, pkt);
1305                         break;
1306                 case RESPST_CHK_LENGTH:
1307                         state = check_length(qp, pkt);
1308                         break;
1309                 case RESPST_CHK_RKEY:
1310                         state = check_rkey(qp, pkt);
1311                         break;
1312                 case RESPST_EXECUTE:
1313                         state = execute(qp, pkt);
1314                         break;
1315                 case RESPST_COMPLETE:
1316                         state = do_complete(qp, pkt);
1317                         break;
1318                 case RESPST_READ_REPLY:
1319                         state = read_reply(qp, pkt);
1320                         break;
1321                 case RESPST_ACKNOWLEDGE:
1322                         state = acknowledge(qp, pkt);
1323                         break;
1324                 case RESPST_CLEANUP:
1325                         state = cleanup(qp, pkt);
1326                         break;
1327                 case RESPST_DUPLICATE_REQUEST:
1328                         state = duplicate_request(qp, pkt);
1329                         break;
1330                 case RESPST_ERR_PSN_OUT_OF_SEQ:
1331                         /* RC only - Class B. Drop packet. */
1332                         send_ack(qp, pkt, AETH_NAK_PSN_SEQ_ERROR, qp->resp.psn);
1333                         state = RESPST_CLEANUP;
1334                         break;
1335
1336                 case RESPST_ERR_TOO_MANY_RDMA_ATM_REQ:
1337                 case RESPST_ERR_MISSING_OPCODE_FIRST:
1338                 case RESPST_ERR_MISSING_OPCODE_LAST_C:
1339                 case RESPST_ERR_UNSUPPORTED_OPCODE:
1340                 case RESPST_ERR_MISALIGNED_ATOMIC:
1341                         /* RC Only - Class C. */
1342                         do_class_ac_error(qp, AETH_NAK_INVALID_REQ,
1343                                           IB_WC_REM_INV_REQ_ERR);
1344                         state = RESPST_COMPLETE;
1345                         break;
1346
1347                 case RESPST_ERR_MISSING_OPCODE_LAST_D1E:
1348                         state = do_class_d1e_error(qp);
1349                         break;
1350                 case RESPST_ERR_RNR:
1351                         if (qp_type(qp) == IB_QPT_RC) {
1352                                 rxe_counter_inc(rxe, RXE_CNT_SND_RNR);
1353                                 /* RC - class B */
1354                                 send_ack(qp, pkt, AETH_RNR_NAK |
1355                                          (~AETH_TYPE_MASK &
1356                                          qp->attr.min_rnr_timer),
1357                                          pkt->psn);
1358                         } else {
1359                                 /* UD/UC - class D */
1360                                 qp->resp.drop_msg = 1;
1361                         }
1362                         state = RESPST_CLEANUP;
1363                         break;
1364
1365                 case RESPST_ERR_RKEY_VIOLATION:
1366                         if (qp_type(qp) == IB_QPT_RC) {
1367                                 /* Class C */
1368                                 do_class_ac_error(qp, AETH_NAK_REM_ACC_ERR,
1369                                                   IB_WC_REM_ACCESS_ERR);
1370                                 state = RESPST_COMPLETE;
1371                         } else {
1372                                 qp->resp.drop_msg = 1;
1373                                 if (qp->srq) {
1374                                         /* UC/SRQ Class D */
1375                                         qp->resp.status = IB_WC_REM_ACCESS_ERR;
1376                                         state = RESPST_COMPLETE;
1377                                 } else {
1378                                         /* UC/non-SRQ Class E. */
1379                                         state = RESPST_CLEANUP;
1380                                 }
1381                         }
1382                         break;
1383
1384                 case RESPST_ERR_INVALIDATE_RKEY:
1385                         /* RC - Class J. */
1386                         qp->resp.goto_error = 1;
1387                         qp->resp.status = IB_WC_REM_INV_REQ_ERR;
1388                         state = RESPST_COMPLETE;
1389                         break;
1390
1391                 case RESPST_ERR_LENGTH:
1392                         if (qp_type(qp) == IB_QPT_RC) {
1393                                 /* Class C */
1394                                 do_class_ac_error(qp, AETH_NAK_INVALID_REQ,
1395                                                   IB_WC_REM_INV_REQ_ERR);
1396                                 state = RESPST_COMPLETE;
1397                         } else if (qp->srq) {
1398                                 /* UC/UD - class E */
1399                                 qp->resp.status = IB_WC_REM_INV_REQ_ERR;
1400                                 state = RESPST_COMPLETE;
1401                         } else {
1402                                 /* UC/UD - class D */
1403                                 qp->resp.drop_msg = 1;
1404                                 state = RESPST_CLEANUP;
1405                         }
1406                         break;
1407
1408                 case RESPST_ERR_MALFORMED_WQE:
1409                         /* All, Class A. */
1410                         do_class_ac_error(qp, AETH_NAK_REM_OP_ERR,
1411                                           IB_WC_LOC_QP_OP_ERR);
1412                         state = RESPST_COMPLETE;
1413                         break;
1414
1415                 case RESPST_ERR_CQ_OVERFLOW:
1416                         /* All - Class G */
1417                         state = RESPST_ERROR;
1418                         break;
1419
1420                 case RESPST_DONE:
1421                         if (qp->resp.goto_error) {
1422                                 state = RESPST_ERROR;
1423                                 break;
1424                         }
1425
1426                         goto done;
1427
1428                 case RESPST_EXIT:
1429                         if (qp->resp.goto_error) {
1430                                 state = RESPST_ERROR;
1431                                 break;
1432                         }
1433
1434                         goto exit;
1435
1436                 case RESPST_RESET:
1437                         rxe_drain_req_pkts(qp, false);
1438                         qp->resp.wqe = NULL;
1439                         goto exit;
1440
1441                 case RESPST_ERROR:
1442                         qp->resp.goto_error = 0;
1443                         pr_warn("qp#%d moved to error state\n", qp_num(qp));
1444                         rxe_qp_error(qp);
1445                         goto exit;
1446
1447                 default:
1448                         WARN_ON_ONCE(1);
1449                 }
1450         }
1451
1452 exit:
1453         ret = -EAGAIN;
1454 done:
1455         rxe_put(qp);
1456         return ret;
1457 }