drivers/infiniband/sw/rxe/rxe_resp.c

   1 /*
   2  * Copyright (c) 2016 Mellanox Technologies Ltd. All rights reserved.
   3  * Copyright (c) 2015 System Fabric Works, Inc. All rights reserved.
   4  *
   5  * This software is available to you under a choice of one of two
   6  * licenses.  You may choose to be licensed under the terms of the GNU
   7  * General Public License (GPL) Version 2, available from the file
   8  * COPYING in the main directory of this source tree, or the
   9  * OpenIB.org BSD license below:
  10  *
  11  *     Redistribution and use in source and binary forms, with or
  12  *     without modification, are permitted provided that the following
  13  *     conditions are met:
  14  *
  15  *      - Redistributions of source code must retain the above
  16  *        copyright notice, this list of conditions and the following
  17  *        disclaimer.
  18  *
  19  *      - Redistributions in binary form must reproduce the above
  20  *        copyright notice, this list of conditions and the following
  21  *        disclaimer in the documentation and/or other materials
  22  *        provided with the distribution.
  23  *
  24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  31  * SOFTWARE.
  32  */
  33
  34 #include <linux/skbuff.h>
  35
  36 #include "rxe.h"
  37 #include "rxe_loc.h"
  38 #include "rxe_queue.h"
  39
  40 enum resp_states {
  41         RESPST_NONE,
  42         RESPST_GET_REQ,
  43         RESPST_CHK_PSN,
  44         RESPST_CHK_OP_SEQ,
  45         RESPST_CHK_OP_VALID,
  46         RESPST_CHK_RESOURCE,
  47         RESPST_CHK_LENGTH,
  48         RESPST_CHK_RKEY,
  49         RESPST_EXECUTE,
  50         RESPST_READ_REPLY,
  51         RESPST_COMPLETE,
  52         RESPST_ACKNOWLEDGE,
  53         RESPST_CLEANUP,
  54         RESPST_DUPLICATE_REQUEST,
  55         RESPST_ERR_MALFORMED_WQE,
  56         RESPST_ERR_UNSUPPORTED_OPCODE,
  57         RESPST_ERR_MISALIGNED_ATOMIC,
  58         RESPST_ERR_PSN_OUT_OF_SEQ,
  59         RESPST_ERR_MISSING_OPCODE_FIRST,
  60         RESPST_ERR_MISSING_OPCODE_LAST_C,
  61         RESPST_ERR_MISSING_OPCODE_LAST_D1E,
  62         RESPST_ERR_TOO_MANY_RDMA_ATM_REQ,
  63         RESPST_ERR_RNR,
  64         RESPST_ERR_RKEY_VIOLATION,
  65         RESPST_ERR_LENGTH,
  66         RESPST_ERR_CQ_OVERFLOW,
  67         RESPST_ERROR,
  68         RESPST_RESET,
  69         RESPST_DONE,
  70         RESPST_EXIT,
  71 };
  72
  73 static char *resp_state_name[] = {
  74         [RESPST_NONE]                           = "NONE",
  75         [RESPST_GET_REQ]                        = "GET_REQ",
  76         [RESPST_CHK_PSN]                        = "CHK_PSN",
  77         [RESPST_CHK_OP_SEQ]                     = "CHK_OP_SEQ",
  78         [RESPST_CHK_OP_VALID]                   = "CHK_OP_VALID",
  79         [RESPST_CHK_RESOURCE]                   = "CHK_RESOURCE",
  80         [RESPST_CHK_LENGTH]                     = "CHK_LENGTH",
  81         [RESPST_CHK_RKEY]                       = "CHK_RKEY",
  82         [RESPST_EXECUTE]                        = "EXECUTE",
  83         [RESPST_READ_REPLY]                     = "READ_REPLY",
  84         [RESPST_COMPLETE]                       = "COMPLETE",
  85         [RESPST_ACKNOWLEDGE]                    = "ACKNOWLEDGE",
  86         [RESPST_CLEANUP]                        = "CLEANUP",
  87         [RESPST_DUPLICATE_REQUEST]              = "DUPLICATE_REQUEST",
  88         [RESPST_ERR_MALFORMED_WQE]              = "ERR_MALFORMED_WQE",
  89         [RESPST_ERR_UNSUPPORTED_OPCODE]         = "ERR_UNSUPPORTED_OPCODE",
  90         [RESPST_ERR_MISALIGNED_ATOMIC]          = "ERR_MISALIGNED_ATOMIC",
  91         [RESPST_ERR_PSN_OUT_OF_SEQ]             = "ERR_PSN_OUT_OF_SEQ",
  92         [RESPST_ERR_MISSING_OPCODE_FIRST]       = "ERR_MISSING_OPCODE_FIRST",
  93         [RESPST_ERR_MISSING_OPCODE_LAST_C]      = "ERR_MISSING_OPCODE_LAST_C",
  94         [RESPST_ERR_MISSING_OPCODE_LAST_D1E]    = "ERR_MISSING_OPCODE_LAST_D1E",
  95         [RESPST_ERR_TOO_MANY_RDMA_ATM_REQ]      = "ERR_TOO_MANY_RDMA_ATM_REQ",
  96         [RESPST_ERR_RNR]                        = "ERR_RNR",
  97         [RESPST_ERR_RKEY_VIOLATION]             = "ERR_RKEY_VIOLATION",
  98         [RESPST_ERR_LENGTH]                     = "ERR_LENGTH",
  99         [RESPST_ERR_CQ_OVERFLOW]                = "ERR_CQ_OVERFLOW",
 100         [RESPST_ERROR]                          = "ERROR",
 101         [RESPST_RESET]                          = "RESET",
 102         [RESPST_DONE]                           = "DONE",
 103         [RESPST_EXIT]                           = "EXIT",
 104 };
 105
 106 /* rxe_recv calls here to add a request packet to the input queue */
 107 void rxe_resp_queue_pkt(struct rxe_dev *rxe, struct rxe_qp *qp,
 108                         struct sk_buff *skb)
 109 {
 110         int must_sched;
 111         struct rxe_pkt_info *pkt = SKB_TO_PKT(skb);
 112
 113         skb_queue_tail(&qp->req_pkts, skb);
 114
 115         must_sched = (pkt->opcode == IB_OPCODE_RC_RDMA_READ_REQUEST) ||
 116                         (skb_queue_len(&qp->req_pkts) > 1);
 117
 118         rxe_run_task(&qp->resp.task, must_sched);
 119 }
 120
 121 static inline enum resp_states get_req(struct rxe_qp *qp,
 122                                        struct rxe_pkt_info **pkt_p)
 123 {
 124         struct sk_buff *skb;
 125
 126         if (qp->resp.state == QP_STATE_ERROR) {
 127                 skb = skb_dequeue(&qp->req_pkts);
 128                 if (skb) {
 129                         /* drain request packet queue */
 130                         rxe_drop_ref(qp);
 131                         kfree_skb(skb);
 132                         return RESPST_GET_REQ;
 133                 }
 134
 135                 /* go drain recv wr queue */
 136                 return RESPST_CHK_RESOURCE;
 137         }
 138
 139         skb = skb_peek(&qp->req_pkts);
 140         if (!skb)
 141                 return RESPST_EXIT;
 142
 143         *pkt_p = SKB_TO_PKT(skb);
 144
 145         return (qp->resp.res) ? RESPST_READ_REPLY : RESPST_CHK_PSN;
 146 }
 147
 148 static enum resp_states check_psn(struct rxe_qp *qp,
 149                                   struct rxe_pkt_info *pkt)
 150 {
 151         int diff = psn_compare(pkt->psn, qp->resp.psn);
 152         struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
 153
 154         switch (qp_type(qp)) {
 155         case IB_QPT_RC:
 156                 if (diff > 0) {
 157                         if (qp->resp.sent_psn_nak)
 158                                 return RESPST_CLEANUP;
 159
 160                         qp->resp.sent_psn_nak = 1;
 161                         rxe_counter_inc(rxe, RXE_CNT_OUT_OF_SEQ_REQ);
 162                         return RESPST_ERR_PSN_OUT_OF_SEQ;
 163
 164                 } else if (diff < 0) {
 165                         rxe_counter_inc(rxe, RXE_CNT_DUP_REQ);
 166                         return RESPST_DUPLICATE_REQUEST;
 167                 }
 168
 169                 if (qp->resp.sent_psn_nak)
 170                         qp->resp.sent_psn_nak = 0;
 171
 172                 break;
 173
 174         case IB_QPT_UC:
 175                 if (qp->resp.drop_msg || diff != 0) {
 176                         if (pkt->mask & RXE_START_MASK) {
 177                                 qp->resp.drop_msg = 0;
 178                                 return RESPST_CHK_OP_SEQ;
 179                         }
 180
 181                         qp->resp.drop_msg = 1;
 182                         return RESPST_CLEANUP;
 183                 }
 184                 break;
 185         default:
 186                 break;
 187         }
 188
 189         return RESPST_CHK_OP_SEQ;
 190 }
 191
 192 static enum resp_states check_op_seq(struct rxe_qp *qp,
 193                                      struct rxe_pkt_info *pkt)
 194 {
 195         switch (qp_type(qp)) {
 196         case IB_QPT_RC:
 197                 switch (qp->resp.opcode) {
 198                 case IB_OPCODE_RC_SEND_FIRST:
 199                 case IB_OPCODE_RC_SEND_MIDDLE:
 200                         switch (pkt->opcode) {
 201                         case IB_OPCODE_RC_SEND_MIDDLE:
 202                         case IB_OPCODE_RC_SEND_LAST:
 203                         case IB_OPCODE_RC_SEND_LAST_WITH_IMMEDIATE:
 204                         case IB_OPCODE_RC_SEND_LAST_WITH_INVALIDATE:
 205                                 return RESPST_CHK_OP_VALID;
 206                         default:
 207                                 return RESPST_ERR_MISSING_OPCODE_LAST_C;
 208                         }
 209
 210                 case IB_OPCODE_RC_RDMA_WRITE_FIRST:
 211                 case IB_OPCODE_RC_RDMA_WRITE_MIDDLE:
 212                         switch (pkt->opcode) {
 213                         case IB_OPCODE_RC_RDMA_WRITE_MIDDLE:
 214                         case IB_OPCODE_RC_RDMA_WRITE_LAST:
 215                         case IB_OPCODE_RC_RDMA_WRITE_LAST_WITH_IMMEDIATE:
 216                                 return RESPST_CHK_OP_VALID;
 217                         default:
 218                                 return RESPST_ERR_MISSING_OPCODE_LAST_C;
 219                         }
 220
 221                 default:
 222                         switch (pkt->opcode) {
 223                         case IB_OPCODE_RC_SEND_MIDDLE:
 224                         case IB_OPCODE_RC_SEND_LAST:
 225                         case IB_OPCODE_RC_SEND_LAST_WITH_IMMEDIATE:
 226                         case IB_OPCODE_RC_SEND_LAST_WITH_INVALIDATE:
 227                         case IB_OPCODE_RC_RDMA_WRITE_MIDDLE:
 228                         case IB_OPCODE_RC_RDMA_WRITE_LAST:
 229                         case IB_OPCODE_RC_RDMA_WRITE_LAST_WITH_IMMEDIATE:
 230                                 return RESPST_ERR_MISSING_OPCODE_FIRST;
 231                         default:
 232                                 return RESPST_CHK_OP_VALID;
 233                         }
 234                 }
 235                 break;
 236
 237         case IB_QPT_UC:
 238                 switch (qp->resp.opcode) {
 239                 case IB_OPCODE_UC_SEND_FIRST:
 240                 case IB_OPCODE_UC_SEND_MIDDLE:
 241                         switch (pkt->opcode) {
 242                         case IB_OPCODE_UC_SEND_MIDDLE:
 243                         case IB_OPCODE_UC_SEND_LAST:
 244                         case IB_OPCODE_UC_SEND_LAST_WITH_IMMEDIATE:
 245                                 return RESPST_CHK_OP_VALID;
 246                         default:
 247                                 return RESPST_ERR_MISSING_OPCODE_LAST_D1E;
 248                         }
 249
 250                 case IB_OPCODE_UC_RDMA_WRITE_FIRST:
 251                 case IB_OPCODE_UC_RDMA_WRITE_MIDDLE:
 252                         switch (pkt->opcode) {
 253                         case IB_OPCODE_UC_RDMA_WRITE_MIDDLE:
 254                         case IB_OPCODE_UC_RDMA_WRITE_LAST:
 255                         case IB_OPCODE_UC_RDMA_WRITE_LAST_WITH_IMMEDIATE:
 256                                 return RESPST_CHK_OP_VALID;
 257                         default:
 258                                 return RESPST_ERR_MISSING_OPCODE_LAST_D1E;
 259                         }
 260
 261                 default:
 262                         switch (pkt->opcode) {
 263                         case IB_OPCODE_UC_SEND_MIDDLE:
 264                         case IB_OPCODE_UC_SEND_LAST:
 265                         case IB_OPCODE_UC_SEND_LAST_WITH_IMMEDIATE:
 266                         case IB_OPCODE_UC_RDMA_WRITE_MIDDLE:
 267                         case IB_OPCODE_UC_RDMA_WRITE_LAST:
 268                         case IB_OPCODE_UC_RDMA_WRITE_LAST_WITH_IMMEDIATE:
 269                                 qp->resp.drop_msg = 1;
 270                                 return RESPST_CLEANUP;
 271                         default:
 272                                 return RESPST_CHK_OP_VALID;
 273                         }
 274                 }
 275                 break;
 276
 277         default:
 278                 return RESPST_CHK_OP_VALID;
 279         }
 280 }
 281
 282 static enum resp_states check_op_valid(struct rxe_qp *qp,
 283                                        struct rxe_pkt_info *pkt)
 284 {
 285         switch (qp_type(qp)) {
 286         case IB_QPT_RC:
 287                 if (((pkt->mask & RXE_READ_MASK) &&
 288                      !(qp->attr.qp_access_flags & IB_ACCESS_REMOTE_READ)) ||
 289                     ((pkt->mask & RXE_WRITE_MASK) &&
 290                      !(qp->attr.qp_access_flags & IB_ACCESS_REMOTE_WRITE)) ||
 291                     ((pkt->mask & RXE_ATOMIC_MASK) &&
 292                      !(qp->attr.qp_access_flags & IB_ACCESS_REMOTE_ATOMIC))) {
 293                         return RESPST_ERR_UNSUPPORTED_OPCODE;
 294                 }
 295
 296                 break;
 297
 298         case IB_QPT_UC:
 299                 if ((pkt->mask & RXE_WRITE_MASK) &&
 300                     !(qp->attr.qp_access_flags & IB_ACCESS_REMOTE_WRITE)) {
 301                         qp->resp.drop_msg = 1;
 302                         return RESPST_CLEANUP;
 303                 }
 304
 305                 break;
 306
 307         case IB_QPT_UD:
 308         case IB_QPT_SMI:
 309         case IB_QPT_GSI:
 310                 break;
 311
 312         default:
 313                 WARN_ON_ONCE(1);
 314                 break;
 315         }
 316
 317         return RESPST_CHK_RESOURCE;
 318 }
 319
 320 static enum resp_states get_srq_wqe(struct rxe_qp *qp)
 321 {
 322         struct rxe_srq *srq = qp->srq;
 323         struct rxe_queue *q = srq->rq.queue;
 324         struct rxe_recv_wqe *wqe;
 325         struct ib_event ev;
 326
 327         if (srq->error)
 328                 return RESPST_ERR_RNR;
 329
 330         spin_lock_bh(&srq->rq.consumer_lock);
 331
 332         wqe = queue_head(q);
 333         if (!wqe) {
 334                 spin_unlock_bh(&srq->rq.consumer_lock);
 335                 return RESPST_ERR_RNR;
 336         }
 337
 338         /* note kernel and user space recv wqes have same size */
 339         memcpy(&qp->resp.srq_wqe, wqe, sizeof(qp->resp.srq_wqe));
 340
 341         qp->resp.wqe = &qp->resp.srq_wqe.wqe;
 342         advance_consumer(q);
 343
 344         if (srq->limit && srq->ibsrq.event_handler &&
 345             (queue_count(q) < srq->limit)) {
 346                 srq->limit = 0;
 347                 goto event;
 348         }
 349
 350         spin_unlock_bh(&srq->rq.consumer_lock);
 351         return RESPST_CHK_LENGTH;
 352
 353 event:
 354         spin_unlock_bh(&srq->rq.consumer_lock);
 355         ev.device = qp->ibqp.device;
 356         ev.element.srq = qp->ibqp.srq;
 357         ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
 358         srq->ibsrq.event_handler(&ev, srq->ibsrq.srq_context);
 359         return RESPST_CHK_LENGTH;
 360 }
 361
 362 static enum resp_states check_resource(struct rxe_qp *qp,
 363                                        struct rxe_pkt_info *pkt)
 364 {
 365         struct rxe_srq *srq = qp->srq;
 366
 367         if (qp->resp.state == QP_STATE_ERROR) {
 368                 if (qp->resp.wqe) {
 369                         qp->resp.status = IB_WC_WR_FLUSH_ERR;
 370                         return RESPST_COMPLETE;
 371                 } else if (!srq) {
 372                         qp->resp.wqe = queue_head(qp->rq.queue);
 373                         if (qp->resp.wqe) {
 374                                 qp->resp.status = IB_WC_WR_FLUSH_ERR;
 375                                 return RESPST_COMPLETE;
 376                         } else {
 377                                 return RESPST_EXIT;
 378                         }
 379                 } else {
 380                         return RESPST_EXIT;
 381                 }
 382         }
 383
 384         if (pkt->mask & RXE_READ_OR_ATOMIC) {
 385                 /* it is the requesters job to not send
 386                  * too many read/atomic ops, we just
 387                  * recycle the responder resource queue
 388                  */
 389                 if (likely(qp->attr.max_dest_rd_atomic > 0))
 390                         return RESPST_CHK_LENGTH;
 391                 else
 392                         return RESPST_ERR_TOO_MANY_RDMA_ATM_REQ;
 393         }
 394
 395         if (pkt->mask & RXE_RWR_MASK) {
 396                 if (srq)
 397                         return get_srq_wqe(qp);
 398
 399                 qp->resp.wqe = queue_head(qp->rq.queue);
 400                 return (qp->resp.wqe) ? RESPST_CHK_LENGTH : RESPST_ERR_RNR;
 401         }
 402
 403         return RESPST_CHK_LENGTH;
 404 }
 405
 406 static enum resp_states check_length(struct rxe_qp *qp,
 407                                      struct rxe_pkt_info *pkt)
 408 {
 409         switch (qp_type(qp)) {
 410         case IB_QPT_RC:
 411                 return RESPST_CHK_RKEY;
 412
 413         case IB_QPT_UC:
 414                 return RESPST_CHK_RKEY;
 415
 416         default:
 417                 return RESPST_CHK_RKEY;
 418         }
 419 }
 420
 421 static enum resp_states check_rkey(struct rxe_qp *qp,
 422                                    struct rxe_pkt_info *pkt)
 423 {
 424         struct rxe_mem *mem = NULL;
 425         u64 va;
 426         u32 rkey;
 427         u32 resid;
 428         u32 pktlen;
 429         int mtu = qp->mtu;
 430         enum resp_states state;
 431         int access;
 432
 433         if (pkt->mask & (RXE_READ_MASK | RXE_WRITE_MASK)) {
 434                 if (pkt->mask & RXE_RETH_MASK) {
 435                         qp->resp.va = reth_va(pkt);
 436                         qp->resp.rkey = reth_rkey(pkt);
 437                         qp->resp.resid = reth_len(pkt);
 438                 }
 439                 access = (pkt->mask & RXE_READ_MASK) ? IB_ACCESS_REMOTE_READ
 440                                                      : IB_ACCESS_REMOTE_WRITE;
 441         } else if (pkt->mask & RXE_ATOMIC_MASK) {
 442                 qp->resp.va = atmeth_va(pkt);
 443                 qp->resp.rkey = atmeth_rkey(pkt);
 444                 qp->resp.resid = sizeof(u64);
 445                 access = IB_ACCESS_REMOTE_ATOMIC;
 446         } else {
 447                 return RESPST_EXECUTE;
 448         }
 449
 450         /* A zero-byte op is not required to set an addr or rkey. */
 451         if ((pkt->mask & (RXE_READ_MASK | RXE_WRITE_OR_SEND)) &&
 452             (pkt->mask & RXE_RETH_MASK) &&
 453             reth_len(pkt) == 0) {
 454                 return RESPST_EXECUTE;
 455         }
 456
 457         va      = qp->resp.va;
 458         rkey    = qp->resp.rkey;
 459         resid   = qp->resp.resid;
 460         pktlen  = payload_size(pkt);
 461
 462         mem = lookup_mem(qp->pd, access, rkey, lookup_remote);
 463         if (!mem) {
 464                 state = RESPST_ERR_RKEY_VIOLATION;
 465                 goto err;
 466         }
 467
 468         if (unlikely(mem->state == RXE_MEM_STATE_FREE)) {
 469                 state = RESPST_ERR_RKEY_VIOLATION;
 470                 goto err;
 471         }
 472
 473         if (mem_check_range(mem, va, resid)) {
 474                 state = RESPST_ERR_RKEY_VIOLATION;
 475                 goto err;
 476         }
 477
 478         if (pkt->mask & RXE_WRITE_MASK)  {
 479                 if (resid > mtu) {
 480                         if (pktlen != mtu || bth_pad(pkt)) {
 481                                 state = RESPST_ERR_LENGTH;
 482                                 goto err;
 483                         }
 484                 } else {
 485                         if (pktlen != resid) {
 486                                 state = RESPST_ERR_LENGTH;
 487                                 goto err;
 488                         }
 489                         if ((bth_pad(pkt) != (0x3 & (-resid)))) {
 490                                 /* This case may not be exactly that
 491                                  * but nothing else fits.
 492                                  */
 493                                 state = RESPST_ERR_LENGTH;
 494                                 goto err;
 495                         }
 496                 }
 497         }
 498
 499         WARN_ON_ONCE(qp->resp.mr);
 500
 501         qp->resp.mr = mem;
 502         return RESPST_EXECUTE;
 503
 504 err:
 505         if (mem)
 506                 rxe_drop_ref(mem);
 507         return state;
 508 }
 509
 510 static enum resp_states send_data_in(struct rxe_qp *qp, void *data_addr,
 511                                      int data_len)
 512 {
 513         int err;
 514         struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
 515
 516         err = copy_data(rxe, qp->pd, IB_ACCESS_LOCAL_WRITE, &qp->resp.wqe->dma,
 517                         data_addr, data_len, to_mem_obj, NULL);
 518         if (unlikely(err))
 519                 return (err == -ENOSPC) ? RESPST_ERR_LENGTH
 520                                         : RESPST_ERR_MALFORMED_WQE;
 521
 522         return RESPST_NONE;
 523 }
 524
 525 static enum resp_states write_data_in(struct rxe_qp *qp,
 526                                       struct rxe_pkt_info *pkt)
 527 {
 528         enum resp_states rc = RESPST_NONE;
 529         int     err;
 530         int data_len = payload_size(pkt);
 531
 532         err = rxe_mem_copy(qp->resp.mr, qp->resp.va, payload_addr(pkt),
 533                            data_len, to_mem_obj, NULL);
 534         if (err) {
 535                 rc = RESPST_ERR_RKEY_VIOLATION;
 536                 goto out;
 537         }
 538
 539         qp->resp.va += data_len;
 540         qp->resp.resid -= data_len;
 541
 542 out:
 543         return rc;
 544 }
 545
 546 /* Guarantee atomicity of atomic operations at the machine level. */
 547 static DEFINE_SPINLOCK(atomic_ops_lock);
 548
 549 static enum resp_states process_atomic(struct rxe_qp *qp,
 550                                        struct rxe_pkt_info *pkt)
 551 {
 552         u64 iova = atmeth_va(pkt);
 553         u64 *vaddr;
 554         enum resp_states ret;
 555         struct rxe_mem *mr = qp->resp.mr;
 556
 557         if (mr->state != RXE_MEM_STATE_VALID) {
 558                 ret = RESPST_ERR_RKEY_VIOLATION;
 559                 goto out;
 560         }
 561
 562         vaddr = iova_to_vaddr(mr, iova, sizeof(u64));
 563
 564         /* check vaddr is 8 bytes aligned. */
 565         if (!vaddr || (uintptr_t)vaddr & 7) {
 566                 ret = RESPST_ERR_MISALIGNED_ATOMIC;
 567                 goto out;
 568         }
 569
 570         spin_lock_bh(&atomic_ops_lock);
 571
 572         qp->resp.atomic_orig = *vaddr;
 573
 574         if (pkt->opcode == IB_OPCODE_RC_COMPARE_SWAP ||
 575             pkt->opcode == IB_OPCODE_RD_COMPARE_SWAP) {
 576                 if (*vaddr == atmeth_comp(pkt))
 577                         *vaddr = atmeth_swap_add(pkt);
 578         } else {
 579                 *vaddr += atmeth_swap_add(pkt);
 580         }
 581
 582         spin_unlock_bh(&atomic_ops_lock);
 583
 584         ret = RESPST_NONE;
 585 out:
 586         return ret;
 587 }
 588
 589 static struct sk_buff *prepare_ack_packet(struct rxe_qp *qp,
 590                                           struct rxe_pkt_info *pkt,
 591                                           struct rxe_pkt_info *ack,
 592                                           int opcode,
 593                                           int payload,
 594                                           u32 psn,
 595                                           u8 syndrome,
 596                                           u32 *crcp)
 597 {
 598         struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
 599         struct sk_buff *skb;
 600         u32 crc = 0;
 601         u32 *p;
 602         int paylen;
 603         int pad;
 604         int err;
 605
 606         /*
 607          * allocate packet
 608          */
 609         pad = (-payload) & 0x3;
 610         paylen = rxe_opcode[opcode].length + payload + pad + RXE_ICRC_SIZE;
 611
 612         skb = rxe_init_packet(rxe, &qp->pri_av, paylen, ack);
 613         if (!skb)
 614                 return NULL;
 615
 616         ack->qp = qp;
 617         ack->opcode = opcode;
 618         ack->mask = rxe_opcode[opcode].mask;
 619         ack->offset = pkt->offset;
 620         ack->paylen = paylen;
 621
 622         /* fill in bth using the request packet headers */
 623         memcpy(ack->hdr, pkt->hdr, pkt->offset + RXE_BTH_BYTES);
 624
 625         bth_set_opcode(ack, opcode);
 626         bth_set_qpn(ack, qp->attr.dest_qp_num);
 627         bth_set_pad(ack, pad);
 628         bth_set_se(ack, 0);
 629         bth_set_psn(ack, psn);
 630         bth_set_ack(ack, 0);
 631         ack->psn = psn;
 632
 633         if (ack->mask & RXE_AETH_MASK) {
 634                 aeth_set_syn(ack, syndrome);
 635                 aeth_set_msn(ack, qp->resp.msn);
 636         }
 637
 638         if (ack->mask & RXE_ATMACK_MASK)
 639                 atmack_set_orig(ack, qp->resp.atomic_orig);
 640
 641         err = rxe_prepare(rxe, ack, skb, &crc);
 642         if (err) {
 643                 kfree_skb(skb);
 644                 return NULL;
 645         }
 646
 647         if (crcp) {
 648                 /* CRC computation will be continued by the caller */
 649                 *crcp = crc;
 650         } else {
 651                 p = payload_addr(ack) + payload + bth_pad(ack);
 652                 *p = ~crc;
 653         }
 654
 655         return skb;
 656 }
 657
 658 /* RDMA read response. If res is not NULL, then we have a current RDMA request
 659  * being processed or replayed.
 660  */
 661 static enum resp_states read_reply(struct rxe_qp *qp,
 662                                    struct rxe_pkt_info *req_pkt)
 663 {
 664         struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
 665         struct rxe_pkt_info ack_pkt;
 666         struct sk_buff *skb;
 667         int mtu = qp->mtu;
 668         enum resp_states state;
 669         int payload;
 670         int opcode;
 671         int err;
 672         struct resp_res *res = qp->resp.res;
 673         u32 icrc;
 674         u32 *p;
 675
 676         if (!res) {
 677                 /* This is the first time we process that request. Get a
 678                  * resource
 679                  */
 680                 res = &qp->resp.resources[qp->resp.res_head];
 681
 682                 free_rd_atomic_resource(qp, res);
 683                 rxe_advance_resp_resource(qp);
 684
 685                 res->type               = RXE_READ_MASK;
 686
 687                 res->read.va            = qp->resp.va;
 688                 res->read.va_org        = qp->resp.va;
 689
 690                 res->first_psn          = req_pkt->psn;
 691
 692                 if (reth_len(req_pkt)) {
 693                         res->last_psn   = (req_pkt->psn +
 694                                            (reth_len(req_pkt) + mtu - 1) /
 695                                            mtu - 1) & BTH_PSN_MASK;
 696                 } else {
 697                         res->last_psn   = res->first_psn;
 698                 }
 699                 res->cur_psn            = req_pkt->psn;
 700
 701                 res->read.resid         = qp->resp.resid;
 702                 res->read.length        = qp->resp.resid;
 703                 res->read.rkey          = qp->resp.rkey;
 704
 705                 /* note res inherits the reference to mr from qp */
 706                 res->read.mr            = qp->resp.mr;
 707                 qp->resp.mr             = NULL;
 708
 709                 qp->resp.res            = res;
 710                 res->state              = rdatm_res_state_new;
 711         }
 712
 713         if (res->state == rdatm_res_state_new) {
 714                 if (res->read.resid <= mtu)
 715                         opcode = IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY;
 716                 else
 717                         opcode = IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST;
 718         } else {
 719                 if (res->read.resid > mtu)
 720                         opcode = IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE;
 721                 else
 722                         opcode = IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST;
 723         }
 724
 725         res->state = rdatm_res_state_next;
 726
 727         payload = min_t(int, res->read.resid, mtu);
 728
 729         skb = prepare_ack_packet(qp, req_pkt, &ack_pkt, opcode, payload,
 730                                  res->cur_psn, AETH_ACK_UNLIMITED, &icrc);
 731         if (!skb)
 732                 return RESPST_ERR_RNR;
 733
 734         err = rxe_mem_copy(res->read.mr, res->read.va, payload_addr(&ack_pkt),
 735                            payload, from_mem_obj, &icrc);
 736         if (err)
 737                 pr_err("Failed copying memory\n");
 738
 739         p = payload_addr(&ack_pkt) + payload + bth_pad(&ack_pkt);
 740         *p = ~icrc;
 741
 742         err = rxe_xmit_packet(rxe, qp, &ack_pkt, skb);
 743         if (err) {
 744                 pr_err("Failed sending RDMA reply.\n");
 745                 kfree_skb(skb);
 746                 return RESPST_ERR_RNR;
 747         }
 748
 749         res->read.va += payload;
 750         res->read.resid -= payload;
 751         res->cur_psn = (res->cur_psn + 1) & BTH_PSN_MASK;
 752
 753         if (res->read.resid > 0) {
 754                 state = RESPST_DONE;
 755         } else {
 756                 qp->resp.res = NULL;
 757                 qp->resp.opcode = -1;
 758                 if (psn_compare(res->cur_psn, qp->resp.psn) >= 0)
 759                         qp->resp.psn = res->cur_psn;
 760                 state = RESPST_CLEANUP;
 761         }
 762
 763         return state;
 764 }
 765
 766 static void build_rdma_network_hdr(union rdma_network_hdr *hdr,
 767                                    struct rxe_pkt_info *pkt)
 768 {
 769         struct sk_buff *skb = PKT_TO_SKB(pkt);
 770
 771         memset(hdr, 0, sizeof(*hdr));
 772         if (skb->protocol == htons(ETH_P_IP))
 773                 memcpy(&hdr->roce4grh, ip_hdr(skb), sizeof(hdr->roce4grh));
 774         else if (skb->protocol == htons(ETH_P_IPV6))
 775                 memcpy(&hdr->ibgrh, ipv6_hdr(skb), sizeof(hdr->ibgrh));
 776 }
 777
 778 /* Executes a new request. A retried request never reach that function (send
 779  * and writes are discarded, and reads and atomics are retried elsewhere.
 780  */
 781 static enum resp_states execute(struct rxe_qp *qp, struct rxe_pkt_info *pkt)
 782 {
 783         enum resp_states err;
 784
 785         if (pkt->mask & RXE_SEND_MASK) {
 786                 if (qp_type(qp) == IB_QPT_UD ||
 787                     qp_type(qp) == IB_QPT_SMI ||
 788                     qp_type(qp) == IB_QPT_GSI) {
 789                         union rdma_network_hdr hdr;
 790
 791                         build_rdma_network_hdr(&hdr, pkt);
 792
 793                         err = send_data_in(qp, &hdr, sizeof(hdr));
 794                         if (err)
 795                                 return err;
 796                 }
 797                 err = send_data_in(qp, payload_addr(pkt), payload_size(pkt));
 798                 if (err)
 799                         return err;
 800         } else if (pkt->mask & RXE_WRITE_MASK) {
 801                 err = write_data_in(qp, pkt);
 802                 if (err)
 803                         return err;
 804         } else if (pkt->mask & RXE_READ_MASK) {
 805                 /* For RDMA Read we can increment the msn now. See C9-148. */
 806                 qp->resp.msn++;
 807                 return RESPST_READ_REPLY;
 808         } else if (pkt->mask & RXE_ATOMIC_MASK) {
 809                 err = process_atomic(qp, pkt);
 810                 if (err)
 811                         return err;
 812         } else {
 813                 /* Unreachable */
 814                 WARN_ON_ONCE(1);
 815         }
 816
 817         /* next expected psn, read handles this separately */
 818         qp->resp.psn = (pkt->psn + 1) & BTH_PSN_MASK;
 819
 820         qp->resp.opcode = pkt->opcode;
 821         qp->resp.status = IB_WC_SUCCESS;
 822
 823         if (pkt->mask & RXE_COMP_MASK) {
 824                 /* We successfully processed this new request. */
 825                 qp->resp.msn++;
 826                 return RESPST_COMPLETE;
 827         } else if (qp_type(qp) == IB_QPT_RC)
 828                 return RESPST_ACKNOWLEDGE;
 829         else
 830                 return RESPST_CLEANUP;
 831 }
 832
 833 static enum resp_states do_complete(struct rxe_qp *qp,
 834                                     struct rxe_pkt_info *pkt)
 835 {
 836         struct rxe_cqe cqe;
 837         struct ib_wc *wc = &cqe.ibwc;
 838         struct ib_uverbs_wc *uwc = &cqe.uibwc;
 839         struct rxe_recv_wqe *wqe = qp->resp.wqe;
 840
 841         if (unlikely(!wqe))
 842                 return RESPST_CLEANUP;
 843
 844         memset(&cqe, 0, sizeof(cqe));
 845
 846         wc->wr_id               = wqe->wr_id;
 847         wc->status              = qp->resp.status;
 848         wc->qp                  = &qp->ibqp;
 849
 850         /* fields after status are not required for errors */
 851         if (wc->status == IB_WC_SUCCESS) {
 852                 wc->opcode = (pkt->mask & RXE_IMMDT_MASK &&
 853                                 pkt->mask & RXE_WRITE_MASK) ?
 854                                         IB_WC_RECV_RDMA_WITH_IMM : IB_WC_RECV;
 855                 wc->vendor_err = 0;
 856                 wc->byte_len = wqe->dma.length - wqe->dma.resid;
 857
 858                 /* fields after byte_len are different between kernel and user
 859                  * space
 860                  */
 861                 if (qp->rcq->is_user) {
 862                         uwc->wc_flags = IB_WC_GRH;
 863
 864                         if (pkt->mask & RXE_IMMDT_MASK) {
 865                                 uwc->wc_flags |= IB_WC_WITH_IMM;
 866                                 uwc->ex.imm_data =
 867                                         (__u32 __force)immdt_imm(pkt);
 868                         }
 869
 870                         if (pkt->mask & RXE_IETH_MASK) {
 871                                 uwc->wc_flags |= IB_WC_WITH_INVALIDATE;
 872                                 uwc->ex.invalidate_rkey = ieth_rkey(pkt);
 873                         }
 874
 875                         uwc->qp_num             = qp->ibqp.qp_num;
 876
 877                         if (pkt->mask & RXE_DETH_MASK)
 878                                 uwc->src_qp = deth_sqp(pkt);
 879
 880                         uwc->port_num           = qp->attr.port_num;
 881                 } else {
 882                         struct sk_buff *skb = PKT_TO_SKB(pkt);
 883
 884                         wc->wc_flags = IB_WC_GRH | IB_WC_WITH_NETWORK_HDR_TYPE;
 885                         if (skb->protocol == htons(ETH_P_IP))
 886                                 wc->network_hdr_type = RDMA_NETWORK_IPV4;
 887                         else
 888                                 wc->network_hdr_type = RDMA_NETWORK_IPV6;
 889
 890                         if (pkt->mask & RXE_IMMDT_MASK) {
 891                                 wc->wc_flags |= IB_WC_WITH_IMM;
 892                                 wc->ex.imm_data = immdt_imm(pkt);
 893                         }
 894
 895                         if (pkt->mask & RXE_IETH_MASK) {
 896                                 struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
 897                                 struct rxe_mem *rmr;
 898
 899                                 wc->wc_flags |= IB_WC_WITH_INVALIDATE;
 900                                 wc->ex.invalidate_rkey = ieth_rkey(pkt);
 901
 902                                 rmr = rxe_pool_get_index(&rxe->mr_pool,
 903                                                          wc->ex.invalidate_rkey >> 8);
 904                                 if (unlikely(!rmr)) {
 905                                         pr_err("Bad rkey %#x invalidation\n",
 906                                                wc->ex.invalidate_rkey);
 907                                         return RESPST_ERROR;
 908                                 }
 909                                 rmr->state = RXE_MEM_STATE_FREE;
 910                                 rxe_drop_ref(rmr);
 911                         }
 912
 913                         wc->qp                  = &qp->ibqp;
 914
 915                         if (pkt->mask & RXE_DETH_MASK)
 916                                 wc->src_qp = deth_sqp(pkt);
 917
 918                         wc->port_num            = qp->attr.port_num;
 919                 }
 920         }
 921
 922         /* have copy for srq and reference for !srq */
 923         if (!qp->srq)
 924                 advance_consumer(qp->rq.queue);
 925
 926         qp->resp.wqe = NULL;
 927
 928         if (rxe_cq_post(qp->rcq, &cqe, pkt ? bth_se(pkt) : 1))
 929                 return RESPST_ERR_CQ_OVERFLOW;
 930
 931         if (qp->resp.state == QP_STATE_ERROR)
 932                 return RESPST_CHK_RESOURCE;
 933
 934         if (!pkt)
 935                 return RESPST_DONE;
 936         else if (qp_type(qp) == IB_QPT_RC)
 937                 return RESPST_ACKNOWLEDGE;
 938         else
 939                 return RESPST_CLEANUP;
 940 }
 941
 942 static int send_ack(struct rxe_qp *qp, struct rxe_pkt_info *pkt,
 943                     u8 syndrome, u32 psn)
 944 {
 945         int err = 0;
 946         struct rxe_pkt_info ack_pkt;
 947         struct sk_buff *skb;
 948         struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
 949
 950         skb = prepare_ack_packet(qp, pkt, &ack_pkt, IB_OPCODE_RC_ACKNOWLEDGE,
 951                                  0, psn, syndrome, NULL);
 952         if (!skb) {
 953                 err = -ENOMEM;
 954                 goto err1;
 955         }
 956
 957         err = rxe_xmit_packet(rxe, qp, &ack_pkt, skb);
 958         if (err) {
 959                 pr_err_ratelimited("Failed sending ack\n");
 960                 kfree_skb(skb);
 961         }
 962
 963 err1:
 964         return err;
 965 }
 966
 967 static int send_atomic_ack(struct rxe_qp *qp, struct rxe_pkt_info *pkt,
 968                            u8 syndrome)
 969 {
 970         int rc = 0;
 971         struct rxe_pkt_info ack_pkt;
 972         struct sk_buff *skb;
 973         struct sk_buff *skb_copy;
 974         struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
 975         struct resp_res *res;
 976
 977         skb = prepare_ack_packet(qp, pkt, &ack_pkt,
 978                                  IB_OPCODE_RC_ATOMIC_ACKNOWLEDGE, 0, pkt->psn,
 979                                  syndrome, NULL);
 980         if (!skb) {
 981                 rc = -ENOMEM;
 982                 goto out;
 983         }
 984
 985         skb_copy = skb_clone(skb, GFP_ATOMIC);
 986         if (skb_copy)
 987                 rxe_add_ref(qp); /* for the new SKB */
 988         else {
 989                 pr_warn("Could not clone atomic response\n");
 990                 rc = -ENOMEM;
 991                 goto out;
 992         }
 993
 994         res = &qp->resp.resources[qp->resp.res_head];
 995         free_rd_atomic_resource(qp, res);
 996         rxe_advance_resp_resource(qp);
 997
 998         memcpy(SKB_TO_PKT(skb), &ack_pkt, sizeof(ack_pkt));
 999         memset((unsigned char *)SKB_TO_PKT(skb) + sizeof(ack_pkt), 0,
1000                sizeof(skb->cb) - sizeof(ack_pkt));
1001
1002         res->type = RXE_ATOMIC_MASK;
1003         res->atomic.skb = skb;
1004         res->first_psn = ack_pkt.psn;
1005         res->last_psn  = ack_pkt.psn;
1006         res->cur_psn   = ack_pkt.psn;
1007
1008         rc = rxe_xmit_packet(rxe, qp, &ack_pkt, skb_copy);
1009         if (rc) {
1010                 pr_err_ratelimited("Failed sending ack\n");
1011                 rxe_drop_ref(qp);
1012                 kfree_skb(skb_copy);
1013         }
1014
1015 out:
1016         return rc;
1017 }
1018
1019 static enum resp_states acknowledge(struct rxe_qp *qp,
1020                                     struct rxe_pkt_info *pkt)
1021 {
1022         if (qp_type(qp) != IB_QPT_RC)
1023                 return RESPST_CLEANUP;
1024
1025         if (qp->resp.aeth_syndrome != AETH_ACK_UNLIMITED)
1026                 send_ack(qp, pkt, qp->resp.aeth_syndrome, pkt->psn);
1027         else if (pkt->mask & RXE_ATOMIC_MASK)
1028                 send_atomic_ack(qp, pkt, AETH_ACK_UNLIMITED);
1029         else if (bth_ack(pkt))
1030                 send_ack(qp, pkt, AETH_ACK_UNLIMITED, pkt->psn);
1031
1032         return RESPST_CLEANUP;
1033 }
1034
1035 static enum resp_states cleanup(struct rxe_qp *qp,
1036                                 struct rxe_pkt_info *pkt)
1037 {
1038         struct sk_buff *skb;
1039
1040         if (pkt) {
1041                 skb = skb_dequeue(&qp->req_pkts);
1042                 rxe_drop_ref(qp);
1043                 kfree_skb(skb);
1044         }
1045
1046         if (qp->resp.mr) {
1047                 rxe_drop_ref(qp->resp.mr);
1048                 qp->resp.mr = NULL;
1049         }
1050
1051         return RESPST_DONE;
1052 }
1053
1054 static struct resp_res *find_resource(struct rxe_qp *qp, u32 psn)
1055 {
1056         int i;
1057
1058         for (i = 0; i < qp->attr.max_rd_atomic; i++) {
1059                 struct resp_res *res = &qp->resp.resources[i];
1060
1061                 if (res->type == 0)
1062                         continue;
1063
1064                 if (psn_compare(psn, res->first_psn) >= 0 &&
1065                     psn_compare(psn, res->last_psn) <= 0) {
1066                         return res;
1067                 }
1068         }
1069
1070         return NULL;
1071 }
1072
1073 static enum resp_states duplicate_request(struct rxe_qp *qp,
1074                                           struct rxe_pkt_info *pkt)
1075 {
1076         enum resp_states rc;
1077         u32 prev_psn = (qp->resp.psn - 1) & BTH_PSN_MASK;
1078
1079         if (pkt->mask & RXE_SEND_MASK ||
1080             pkt->mask & RXE_WRITE_MASK) {
1081                 /* SEND. Ack again and cleanup. C9-105. */
1082                 if (bth_ack(pkt))
1083                         send_ack(qp, pkt, AETH_ACK_UNLIMITED, prev_psn);
1084                 rc = RESPST_CLEANUP;
1085                 goto out;
1086         } else if (pkt->mask & RXE_READ_MASK) {
1087                 struct resp_res *res;
1088
1089                 res = find_resource(qp, pkt->psn);
1090                 if (!res) {
1091                         /* Resource not found. Class D error.  Drop the
1092                          * request.
1093                          */
1094                         rc = RESPST_CLEANUP;
1095                         goto out;
1096                 } else {
1097                         /* Ensure this new request is the same as the previous
1098                          * one or a subset of it.
1099                          */
1100                         u64 iova = reth_va(pkt);
1101                         u32 resid = reth_len(pkt);
1102
1103                         if (iova < res->read.va_org ||
1104                             resid > res->read.length ||
1105                             (iova + resid) > (res->read.va_org +
1106                                               res->read.length)) {
1107                                 rc = RESPST_CLEANUP;
1108                                 goto out;
1109                         }
1110
1111                         if (reth_rkey(pkt) != res->read.rkey) {
1112                                 rc = RESPST_CLEANUP;
1113                                 goto out;
1114                         }
1115
1116                         res->cur_psn = pkt->psn;
1117                         res->state = (pkt->psn == res->first_psn) ?
1118                                         rdatm_res_state_new :
1119                                         rdatm_res_state_replay;
1120
1121                         /* Reset the resource, except length. */
1122                         res->read.va_org = iova;
1123                         res->read.va = iova;
1124                         res->read.resid = resid;
1125
1126                         /* Replay the RDMA read reply. */
1127                         qp->resp.res = res;
1128                         rc = RESPST_READ_REPLY;
1129                         goto out;
1130                 }
1131         } else {
1132                 struct resp_res *res;
1133
1134                 /* Find the operation in our list of responder resources. */
1135                 res = find_resource(qp, pkt->psn);
1136                 if (res) {
1137                         struct sk_buff *skb_copy;
1138
1139                         skb_copy = skb_clone(res->atomic.skb, GFP_ATOMIC);
1140                         if (skb_copy) {
1141                                 rxe_add_ref(qp); /* for the new SKB */
1142                         } else {
1143                                 pr_warn("Couldn't clone atomic resp\n");
1144                                 rc = RESPST_CLEANUP;
1145                                 goto out;
1146                         }
1147
1148                         /* Resend the result. */
1149                         rc = rxe_xmit_packet(to_rdev(qp->ibqp.device), qp,
1150                                              pkt, skb_copy);
1151                         if (rc) {
1152                                 pr_err("Failed resending result. This flow is not handled - skb ignored\n");
1153                                 rxe_drop_ref(qp);
1154                                 kfree_skb(skb_copy);
1155                                 rc = RESPST_CLEANUP;
1156                                 goto out;
1157                         }
1158                 }
1159
1160                 /* Resource not found. Class D error. Drop the request. */
1161                 rc = RESPST_CLEANUP;
1162                 goto out;
1163         }
1164 out:
1165         return rc;
1166 }
1167
1168 /* Process a class A or C. Both are treated the same in this implementation. */
1169 static void do_class_ac_error(struct rxe_qp *qp, u8 syndrome,
1170                               enum ib_wc_status status)
1171 {
1172         qp->resp.aeth_syndrome  = syndrome;
1173         qp->resp.status         = status;
1174
1175         /* indicate that we should go through the ERROR state */
1176         qp->resp.goto_error     = 1;
1177 }
1178
1179 static enum resp_states do_class_d1e_error(struct rxe_qp *qp)
1180 {
1181         /* UC */
1182         if (qp->srq) {
1183                 /* Class E */
1184                 qp->resp.drop_msg = 1;
1185                 if (qp->resp.wqe) {
1186                         qp->resp.status = IB_WC_REM_INV_REQ_ERR;
1187                         return RESPST_COMPLETE;
1188                 } else {
1189                         return RESPST_CLEANUP;
1190                 }
1191         } else {
1192                 /* Class D1. This packet may be the start of a
1193                  * new message and could be valid. The previous
1194                  * message is invalid and ignored. reset the
1195                  * recv wr to its original state
1196                  */
1197                 if (qp->resp.wqe) {
1198                         qp->resp.wqe->dma.resid = qp->resp.wqe->dma.length;
1199                         qp->resp.wqe->dma.cur_sge = 0;
1200                         qp->resp.wqe->dma.sge_offset = 0;
1201                         qp->resp.opcode = -1;
1202                 }
1203
1204                 if (qp->resp.mr) {
1205                         rxe_drop_ref(qp->resp.mr);
1206                         qp->resp.mr = NULL;
1207                 }
1208
1209                 return RESPST_CLEANUP;
1210         }
1211 }
1212
1213 void rxe_drain_req_pkts(struct rxe_qp *qp, bool notify)
1214 {
1215         struct sk_buff *skb;
1216
1217         while ((skb = skb_dequeue(&qp->req_pkts))) {
1218                 rxe_drop_ref(qp);
1219                 kfree_skb(skb);
1220         }
1221
1222         if (notify)
1223                 return;
1224
1225         while (!qp->srq && qp->rq.queue && queue_head(qp->rq.queue))
1226                 advance_consumer(qp->rq.queue);
1227 }
1228
1229 int rxe_responder(void *arg)
1230 {
1231         struct rxe_qp *qp = (struct rxe_qp *)arg;
1232         struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
1233         enum resp_states state;
1234         struct rxe_pkt_info *pkt = NULL;
1235         int ret = 0;
1236
1237         rxe_add_ref(qp);
1238
1239         qp->resp.aeth_syndrome = AETH_ACK_UNLIMITED;
1240
1241         if (!qp->valid) {
1242                 ret = -EINVAL;
1243                 goto done;
1244         }
1245
1246         switch (qp->resp.state) {
1247         case QP_STATE_RESET:
1248                 state = RESPST_RESET;
1249                 break;
1250
1251         default:
1252                 state = RESPST_GET_REQ;
1253                 break;
1254         }
1255
1256         while (1) {
1257                 pr_debug("qp#%d state = %s\n", qp_num(qp),
1258                          resp_state_name[state]);
1259                 switch (state) {
1260                 case RESPST_GET_REQ:
1261                         state = get_req(qp, &pkt);
1262                         break;
1263                 case RESPST_CHK_PSN:
1264                         state = check_psn(qp, pkt);
1265                         break;
1266                 case RESPST_CHK_OP_SEQ:
1267                         state = check_op_seq(qp, pkt);
1268                         break;
1269                 case RESPST_CHK_OP_VALID:
1270                         state = check_op_valid(qp, pkt);
1271                         break;
1272                 case RESPST_CHK_RESOURCE:
1273                         state = check_resource(qp, pkt);
1274                         break;
1275                 case RESPST_CHK_LENGTH:
1276                         state = check_length(qp, pkt);
1277                         break;
1278                 case RESPST_CHK_RKEY:
1279                         state = check_rkey(qp, pkt);
1280                         break;
1281                 case RESPST_EXECUTE:
1282                         state = execute(qp, pkt);
1283                         break;
1284                 case RESPST_COMPLETE:
1285                         state = do_complete(qp, pkt);
1286                         break;
1287                 case RESPST_READ_REPLY:
1288                         state = read_reply(qp, pkt);
1289                         break;
1290                 case RESPST_ACKNOWLEDGE:
1291                         state = acknowledge(qp, pkt);
1292                         break;
1293                 case RESPST_CLEANUP:
1294                         state = cleanup(qp, pkt);
1295                         break;
1296                 case RESPST_DUPLICATE_REQUEST:
1297                         state = duplicate_request(qp, pkt);
1298                         break;
1299                 case RESPST_ERR_PSN_OUT_OF_SEQ:
1300                         /* RC only - Class B. Drop packet. */
1301                         send_ack(qp, pkt, AETH_NAK_PSN_SEQ_ERROR, qp->resp.psn);
1302                         state = RESPST_CLEANUP;
1303                         break;
1304
1305                 case RESPST_ERR_TOO_MANY_RDMA_ATM_REQ:
1306                 case RESPST_ERR_MISSING_OPCODE_FIRST:
1307                 case RESPST_ERR_MISSING_OPCODE_LAST_C:
1308                 case RESPST_ERR_UNSUPPORTED_OPCODE:
1309                 case RESPST_ERR_MISALIGNED_ATOMIC:
1310                         /* RC Only - Class C. */
1311                         do_class_ac_error(qp, AETH_NAK_INVALID_REQ,
1312                                           IB_WC_REM_INV_REQ_ERR);
1313                         state = RESPST_COMPLETE;
1314                         break;
1315
1316                 case RESPST_ERR_MISSING_OPCODE_LAST_D1E:
1317                         state = do_class_d1e_error(qp);
1318                         break;
1319                 case RESPST_ERR_RNR:
1320                         if (qp_type(qp) == IB_QPT_RC) {
1321                                 rxe_counter_inc(rxe, RXE_CNT_SND_RNR);
1322                                 /* RC - class B */
1323                                 send_ack(qp, pkt, AETH_RNR_NAK |
1324                                          (~AETH_TYPE_MASK &
1325                                          qp->attr.min_rnr_timer),
1326                                          pkt->psn);
1327                         } else {
1328                                 /* UD/UC - class D */
1329                                 qp->resp.drop_msg = 1;
1330                         }
1331                         state = RESPST_CLEANUP;
1332                         break;
1333
1334                 case RESPST_ERR_RKEY_VIOLATION:
1335                         if (qp_type(qp) == IB_QPT_RC) {
1336                                 /* Class C */
1337                                 do_class_ac_error(qp, AETH_NAK_REM_ACC_ERR,
1338                                                   IB_WC_REM_ACCESS_ERR);
1339                                 state = RESPST_COMPLETE;
1340                         } else {
1341                                 qp->resp.drop_msg = 1;
1342                                 if (qp->srq) {
1343                                         /* UC/SRQ Class D */
1344                                         qp->resp.status = IB_WC_REM_ACCESS_ERR;
1345                                         state = RESPST_COMPLETE;
1346                                 } else {
1347                                         /* UC/non-SRQ Class E. */
1348                                         state = RESPST_CLEANUP;
1349                                 }
1350                         }
1351                         break;
1352
1353                 case RESPST_ERR_LENGTH:
1354                         if (qp_type(qp) == IB_QPT_RC) {
1355                                 /* Class C */
1356                                 do_class_ac_error(qp, AETH_NAK_INVALID_REQ,
1357                                                   IB_WC_REM_INV_REQ_ERR);
1358                                 state = RESPST_COMPLETE;
1359                         } else if (qp->srq) {
1360                                 /* UC/UD - class E */
1361                                 qp->resp.status = IB_WC_REM_INV_REQ_ERR;
1362                                 state = RESPST_COMPLETE;
1363                         } else {
1364                                 /* UC/UD - class D */
1365                                 qp->resp.drop_msg = 1;
1366                                 state = RESPST_CLEANUP;
1367                         }
1368                         break;
1369
1370                 case RESPST_ERR_MALFORMED_WQE:
1371                         /* All, Class A. */
1372                         do_class_ac_error(qp, AETH_NAK_REM_OP_ERR,
1373                                           IB_WC_LOC_QP_OP_ERR);
1374                         state = RESPST_COMPLETE;
1375                         break;
1376
1377                 case RESPST_ERR_CQ_OVERFLOW:
1378                         /* All - Class G */
1379                         state = RESPST_ERROR;
1380                         break;
1381
1382                 case RESPST_DONE:
1383                         if (qp->resp.goto_error) {
1384                                 state = RESPST_ERROR;
1385                                 break;
1386                         }
1387
1388                         goto done;
1389
1390                 case RESPST_EXIT:
1391                         if (qp->resp.goto_error) {
1392                                 state = RESPST_ERROR;
1393                                 break;
1394                         }
1395
1396                         goto exit;
1397
1398                 case RESPST_RESET:
1399                         rxe_drain_req_pkts(qp, false);
1400                         qp->resp.wqe = NULL;
1401                         goto exit;
1402
1403                 case RESPST_ERROR:
1404                         qp->resp.goto_error = 0;
1405                         pr_warn("qp#%d moved to error state\n", qp_num(qp));
1406                         rxe_qp_error(qp);
1407                         goto exit;
1408
1409                 default:
1410                         WARN_ON_ONCE(1);
1411                 }
1412         }
1413
1414 exit:
1415         ret = -EAGAIN;
1416 done:
1417         rxe_drop_ref(qp);
1418         return ret;
1419 }