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                         qp->resp.length = reth_len(pkt);
 439                 }
 440                 access = (pkt->mask & RXE_READ_MASK) ? IB_ACCESS_REMOTE_READ
 441                                                      : IB_ACCESS_REMOTE_WRITE;
 442         } else if (pkt->mask & RXE_ATOMIC_MASK) {
 443                 qp->resp.va = atmeth_va(pkt);
 444                 qp->resp.rkey = atmeth_rkey(pkt);
 445                 qp->resp.resid = sizeof(u64);
 446                 access = IB_ACCESS_REMOTE_ATOMIC;
 447         } else {
 448                 return RESPST_EXECUTE;
 449         }
 450
 451         /* A zero-byte op is not required to set an addr or rkey. */
 452         if ((pkt->mask & (RXE_READ_MASK | RXE_WRITE_OR_SEND)) &&
 453             (pkt->mask & RXE_RETH_MASK) &&
 454             reth_len(pkt) == 0) {
 455                 return RESPST_EXECUTE;
 456         }
 457
 458         va      = qp->resp.va;
 459         rkey    = qp->resp.rkey;
 460         resid   = qp->resp.resid;
 461         pktlen  = payload_size(pkt);
 462
 463         mem = lookup_mem(qp->pd, access, rkey, lookup_remote);
 464         if (!mem) {
 465                 state = RESPST_ERR_RKEY_VIOLATION;
 466                 goto err;
 467         }
 468
 469         if (unlikely(mem->state == RXE_MEM_STATE_FREE)) {
 470                 state = RESPST_ERR_RKEY_VIOLATION;
 471                 goto err;
 472         }
 473
 474         if (mem_check_range(mem, va, resid)) {
 475                 state = RESPST_ERR_RKEY_VIOLATION;
 476                 goto err;
 477         }
 478
 479         if (pkt->mask & RXE_WRITE_MASK)  {
 480                 if (resid > mtu) {
 481                         if (pktlen != mtu || bth_pad(pkt)) {
 482                                 state = RESPST_ERR_LENGTH;
 483                                 goto err;
 484                         }
 485                 } else {
 486                         if (pktlen != resid) {
 487                                 state = RESPST_ERR_LENGTH;
 488                                 goto err;
 489                         }
 490                         if ((bth_pad(pkt) != (0x3 & (-resid)))) {
 491                                 /* This case may not be exactly that
 492                                  * but nothing else fits.
 493                                  */
 494                                 state = RESPST_ERR_LENGTH;
 495                                 goto err;
 496                         }
 497                 }
 498         }
 499
 500         WARN_ON_ONCE(qp->resp.mr);
 501
 502         qp->resp.mr = mem;
 503         return RESPST_EXECUTE;
 504
 505 err:
 506         if (mem)
 507                 rxe_drop_ref(mem);
 508         return state;
 509 }
 510
 511 static enum resp_states send_data_in(struct rxe_qp *qp, void *data_addr,
 512                                      int data_len)
 513 {
 514         int err;
 515
 516         err = copy_data(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                 res->replay             = 0;
 687
 688                 res->read.va            = qp->resp.va;
 689                 res->read.va_org        = qp->resp.va;
 690
 691                 res->first_psn          = req_pkt->psn;
 692
 693                 if (reth_len(req_pkt)) {
 694                         res->last_psn   = (req_pkt->psn +
 695                                            (reth_len(req_pkt) + mtu - 1) /
 696                                            mtu - 1) & BTH_PSN_MASK;
 697                 } else {
 698                         res->last_psn   = res->first_psn;
 699                 }
 700                 res->cur_psn            = req_pkt->psn;
 701
 702                 res->read.resid         = qp->resp.resid;
 703                 res->read.length        = qp->resp.resid;
 704                 res->read.rkey          = qp->resp.rkey;
 705
 706                 /* note res inherits the reference to mr from qp */
 707                 res->read.mr            = qp->resp.mr;
 708                 qp->resp.mr             = NULL;
 709
 710                 qp->resp.res            = res;
 711                 res->state              = rdatm_res_state_new;
 712         }
 713
 714         if (res->state == rdatm_res_state_new) {
 715                 if (res->read.resid <= mtu)
 716                         opcode = IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY;
 717                 else
 718                         opcode = IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST;
 719         } else {
 720                 if (res->read.resid > mtu)
 721                         opcode = IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE;
 722                 else
 723                         opcode = IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST;
 724         }
 725
 726         res->state = rdatm_res_state_next;
 727
 728         payload = min_t(int, res->read.resid, mtu);
 729
 730         skb = prepare_ack_packet(qp, req_pkt, &ack_pkt, opcode, payload,
 731                                  res->cur_psn, AETH_ACK_UNLIMITED, &icrc);
 732         if (!skb)
 733                 return RESPST_ERR_RNR;
 734
 735         err = rxe_mem_copy(res->read.mr, res->read.va, payload_addr(&ack_pkt),
 736                            payload, from_mem_obj, &icrc);
 737         if (err)
 738                 pr_err("Failed copying memory\n");
 739
 740         if (bth_pad(&ack_pkt)) {
 741                 struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
 742                 u8 *pad = payload_addr(&ack_pkt) + payload;
 743
 744                 memset(pad, 0, bth_pad(&ack_pkt));
 745                 icrc = rxe_crc32(rxe, icrc, pad, bth_pad(&ack_pkt));
 746         }
 747         p = payload_addr(&ack_pkt) + payload + bth_pad(&ack_pkt);
 748         *p = ~icrc;
 749
 750         err = rxe_xmit_packet(rxe, qp, &ack_pkt, skb);
 751         if (err) {
 752                 pr_err("Failed sending RDMA reply.\n");
 753                 return RESPST_ERR_RNR;
 754         }
 755
 756         res->read.va += payload;
 757         res->read.resid -= payload;
 758         res->cur_psn = (res->cur_psn + 1) & BTH_PSN_MASK;
 759
 760         if (res->read.resid > 0) {
 761                 state = RESPST_DONE;
 762         } else {
 763                 qp->resp.res = NULL;
 764                 if (!res->replay)
 765                         qp->resp.opcode = -1;
 766                 if (psn_compare(res->cur_psn, qp->resp.psn) >= 0)
 767                         qp->resp.psn = res->cur_psn;
 768                 state = RESPST_CLEANUP;
 769         }
 770
 771         return state;
 772 }
 773
 774 static void build_rdma_network_hdr(union rdma_network_hdr *hdr,
 775                                    struct rxe_pkt_info *pkt)
 776 {
 777         struct sk_buff *skb = PKT_TO_SKB(pkt);
 778
 779         memset(hdr, 0, sizeof(*hdr));
 780         if (skb->protocol == htons(ETH_P_IP))
 781                 memcpy(&hdr->roce4grh, ip_hdr(skb), sizeof(hdr->roce4grh));
 782         else if (skb->protocol == htons(ETH_P_IPV6))
 783                 memcpy(&hdr->ibgrh, ipv6_hdr(skb), sizeof(hdr->ibgrh));
 784 }
 785
 786 /* Executes a new request. A retried request never reach that function (send
 787  * and writes are discarded, and reads and atomics are retried elsewhere.
 788  */
 789 static enum resp_states execute(struct rxe_qp *qp, struct rxe_pkt_info *pkt)
 790 {
 791         enum resp_states err;
 792
 793         if (pkt->mask & RXE_SEND_MASK) {
 794                 if (qp_type(qp) == IB_QPT_UD ||
 795                     qp_type(qp) == IB_QPT_SMI ||
 796                     qp_type(qp) == IB_QPT_GSI) {
 797                         union rdma_network_hdr hdr;
 798
 799                         build_rdma_network_hdr(&hdr, pkt);
 800
 801                         err = send_data_in(qp, &hdr, sizeof(hdr));
 802                         if (err)
 803                                 return err;
 804                 }
 805                 err = send_data_in(qp, payload_addr(pkt), payload_size(pkt));
 806                 if (err)
 807                         return err;
 808         } else if (pkt->mask & RXE_WRITE_MASK) {
 809                 err = write_data_in(qp, pkt);
 810                 if (err)
 811                         return err;
 812         } else if (pkt->mask & RXE_READ_MASK) {
 813                 /* For RDMA Read we can increment the msn now. See C9-148. */
 814                 qp->resp.msn++;
 815                 return RESPST_READ_REPLY;
 816         } else if (pkt->mask & RXE_ATOMIC_MASK) {
 817                 err = process_atomic(qp, pkt);
 818                 if (err)
 819                         return err;
 820         } else {
 821                 /* Unreachable */
 822                 WARN_ON_ONCE(1);
 823         }
 824
 825         /* next expected psn, read handles this separately */
 826         qp->resp.psn = (pkt->psn + 1) & BTH_PSN_MASK;
 827         qp->resp.ack_psn = qp->resp.psn;
 828
 829         qp->resp.opcode = pkt->opcode;
 830         qp->resp.status = IB_WC_SUCCESS;
 831
 832         if (pkt->mask & RXE_COMP_MASK) {
 833                 /* We successfully processed this new request. */
 834                 qp->resp.msn++;
 835                 return RESPST_COMPLETE;
 836         } else if (qp_type(qp) == IB_QPT_RC)
 837                 return RESPST_ACKNOWLEDGE;
 838         else
 839                 return RESPST_CLEANUP;
 840 }
 841
 842 static enum resp_states do_complete(struct rxe_qp *qp,
 843                                     struct rxe_pkt_info *pkt)
 844 {
 845         struct rxe_cqe cqe;
 846         struct ib_wc *wc = &cqe.ibwc;
 847         struct ib_uverbs_wc *uwc = &cqe.uibwc;
 848         struct rxe_recv_wqe *wqe = qp->resp.wqe;
 849
 850         if (unlikely(!wqe))
 851                 return RESPST_CLEANUP;
 852
 853         memset(&cqe, 0, sizeof(cqe));
 854
 855         if (qp->rcq->is_user) {
 856                 uwc->status             = qp->resp.status;
 857                 uwc->qp_num             = qp->ibqp.qp_num;
 858                 uwc->wr_id              = wqe->wr_id;
 859         } else {
 860                 wc->status              = qp->resp.status;
 861                 wc->qp                  = &qp->ibqp;
 862                 wc->wr_id               = wqe->wr_id;
 863         }
 864
 865         if (wc->status == IB_WC_SUCCESS) {
 866                 wc->opcode = (pkt->mask & RXE_IMMDT_MASK &&
 867                                 pkt->mask & RXE_WRITE_MASK) ?
 868                                         IB_WC_RECV_RDMA_WITH_IMM : IB_WC_RECV;
 869                 wc->vendor_err = 0;
 870                 wc->byte_len = (pkt->mask & RXE_IMMDT_MASK &&
 871                                 pkt->mask & RXE_WRITE_MASK) ?
 872                                         qp->resp.length : wqe->dma.length - wqe->dma.resid;
 873
 874                 /* fields after byte_len are different between kernel and user
 875                  * space
 876                  */
 877                 if (qp->rcq->is_user) {
 878                         uwc->wc_flags = IB_WC_GRH;
 879
 880                         if (pkt->mask & RXE_IMMDT_MASK) {
 881                                 uwc->wc_flags |= IB_WC_WITH_IMM;
 882                                 uwc->ex.imm_data = immdt_imm(pkt);
 883                         }
 884
 885                         if (pkt->mask & RXE_IETH_MASK) {
 886                                 uwc->wc_flags |= IB_WC_WITH_INVALIDATE;
 887                                 uwc->ex.invalidate_rkey = ieth_rkey(pkt);
 888                         }
 889
 890                         uwc->qp_num             = qp->ibqp.qp_num;
 891
 892                         if (pkt->mask & RXE_DETH_MASK)
 893                                 uwc->src_qp = deth_sqp(pkt);
 894
 895                         uwc->port_num           = qp->attr.port_num;
 896                 } else {
 897                         struct sk_buff *skb = PKT_TO_SKB(pkt);
 898
 899                         wc->wc_flags = IB_WC_GRH | IB_WC_WITH_NETWORK_HDR_TYPE;
 900                         if (skb->protocol == htons(ETH_P_IP))
 901                                 wc->network_hdr_type = RDMA_NETWORK_IPV4;
 902                         else
 903                                 wc->network_hdr_type = RDMA_NETWORK_IPV6;
 904
 905                         if (is_vlan_dev(skb->dev)) {
 906                                 wc->wc_flags |= IB_WC_WITH_VLAN;
 907                                 wc->vlan_id = vlan_dev_vlan_id(skb->dev);
 908                         }
 909
 910                         if (pkt->mask & RXE_IMMDT_MASK) {
 911                                 wc->wc_flags |= IB_WC_WITH_IMM;
 912                                 wc->ex.imm_data = immdt_imm(pkt);
 913                         }
 914
 915                         if (pkt->mask & RXE_IETH_MASK) {
 916                                 struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
 917                                 struct rxe_mem *rmr;
 918
 919                                 wc->wc_flags |= IB_WC_WITH_INVALIDATE;
 920                                 wc->ex.invalidate_rkey = ieth_rkey(pkt);
 921
 922                                 rmr = rxe_pool_get_index(&rxe->mr_pool,
 923                                                          wc->ex.invalidate_rkey >> 8);
 924                                 if (unlikely(!rmr)) {
 925                                         pr_err("Bad rkey %#x invalidation\n",
 926                                                wc->ex.invalidate_rkey);
 927                                         return RESPST_ERROR;
 928                                 }
 929                                 rmr->state = RXE_MEM_STATE_FREE;
 930                                 rxe_drop_ref(rmr);
 931                         }
 932
 933                         wc->qp                  = &qp->ibqp;
 934
 935                         if (pkt->mask & RXE_DETH_MASK)
 936                                 wc->src_qp = deth_sqp(pkt);
 937
 938                         wc->port_num            = qp->attr.port_num;
 939                 }
 940         }
 941
 942         /* have copy for srq and reference for !srq */
 943         if (!qp->srq)
 944                 advance_consumer(qp->rq.queue);
 945
 946         qp->resp.wqe = NULL;
 947
 948         if (rxe_cq_post(qp->rcq, &cqe, pkt ? bth_se(pkt) : 1))
 949                 return RESPST_ERR_CQ_OVERFLOW;
 950
 951         if (qp->resp.state == QP_STATE_ERROR)
 952                 return RESPST_CHK_RESOURCE;
 953
 954         if (!pkt)
 955                 return RESPST_DONE;
 956         else if (qp_type(qp) == IB_QPT_RC)
 957                 return RESPST_ACKNOWLEDGE;
 958         else
 959                 return RESPST_CLEANUP;
 960 }
 961
 962 static int send_ack(struct rxe_qp *qp, struct rxe_pkt_info *pkt,
 963                     u8 syndrome, u32 psn)
 964 {
 965         int err = 0;
 966         struct rxe_pkt_info ack_pkt;
 967         struct sk_buff *skb;
 968         struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
 969
 970         skb = prepare_ack_packet(qp, pkt, &ack_pkt, IB_OPCODE_RC_ACKNOWLEDGE,
 971                                  0, psn, syndrome, NULL);
 972         if (!skb) {
 973                 err = -ENOMEM;
 974                 goto err1;
 975         }
 976
 977         err = rxe_xmit_packet(rxe, qp, &ack_pkt, skb);
 978         if (err)
 979                 pr_err_ratelimited("Failed sending ack\n");
 980
 981 err1:
 982         return err;
 983 }
 984
 985 static int send_atomic_ack(struct rxe_qp *qp, struct rxe_pkt_info *pkt,
 986                            u8 syndrome)
 987 {
 988         int rc = 0;
 989         struct rxe_pkt_info ack_pkt;
 990         struct sk_buff *skb;
 991         struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
 992         struct resp_res *res;
 993
 994         skb = prepare_ack_packet(qp, pkt, &ack_pkt,
 995                                  IB_OPCODE_RC_ATOMIC_ACKNOWLEDGE, 0, pkt->psn,
 996                                  syndrome, NULL);
 997         if (!skb) {
 998                 rc = -ENOMEM;
 999                 goto out;
1000         }
1001
1002         res = &qp->resp.resources[qp->resp.res_head];
1003         free_rd_atomic_resource(qp, res);
1004         rxe_advance_resp_resource(qp);
1005
1006         memcpy(SKB_TO_PKT(skb), &ack_pkt, sizeof(ack_pkt));
1007         memset((unsigned char *)SKB_TO_PKT(skb) + sizeof(ack_pkt), 0,
1008                sizeof(skb->cb) - sizeof(ack_pkt));
1009
1010         skb_get(skb);
1011         res->type = RXE_ATOMIC_MASK;
1012         res->atomic.skb = skb;
1013         res->first_psn = ack_pkt.psn;
1014         res->last_psn  = ack_pkt.psn;
1015         res->cur_psn   = ack_pkt.psn;
1016
1017         rc = rxe_xmit_packet(rxe, qp, &ack_pkt, skb);
1018         if (rc) {
1019                 pr_err_ratelimited("Failed sending ack\n");
1020                 rxe_drop_ref(qp);
1021         }
1022 out:
1023         return rc;
1024 }
1025
1026 static enum resp_states acknowledge(struct rxe_qp *qp,
1027                                     struct rxe_pkt_info *pkt)
1028 {
1029         if (qp_type(qp) != IB_QPT_RC)
1030                 return RESPST_CLEANUP;
1031
1032         if (qp->resp.aeth_syndrome != AETH_ACK_UNLIMITED)
1033                 send_ack(qp, pkt, qp->resp.aeth_syndrome, pkt->psn);
1034         else if (pkt->mask & RXE_ATOMIC_MASK)
1035                 send_atomic_ack(qp, pkt, AETH_ACK_UNLIMITED);
1036         else if (bth_ack(pkt))
1037                 send_ack(qp, pkt, AETH_ACK_UNLIMITED, pkt->psn);
1038
1039         return RESPST_CLEANUP;
1040 }
1041
1042 static enum resp_states cleanup(struct rxe_qp *qp,
1043                                 struct rxe_pkt_info *pkt)
1044 {
1045         struct sk_buff *skb;
1046
1047         if (pkt) {
1048                 skb = skb_dequeue(&qp->req_pkts);
1049                 rxe_drop_ref(qp);
1050                 kfree_skb(skb);
1051         }
1052
1053         if (qp->resp.mr) {
1054                 rxe_drop_ref(qp->resp.mr);
1055                 qp->resp.mr = NULL;
1056         }
1057
1058         return RESPST_DONE;
1059 }
1060
1061 static struct resp_res *find_resource(struct rxe_qp *qp, u32 psn)
1062 {
1063         int i;
1064
1065         for (i = 0; i < qp->attr.max_dest_rd_atomic; i++) {
1066                 struct resp_res *res = &qp->resp.resources[i];
1067
1068                 if (res->type == 0)
1069                         continue;
1070
1071                 if (psn_compare(psn, res->first_psn) >= 0 &&
1072                     psn_compare(psn, res->last_psn) <= 0) {
1073                         return res;
1074                 }
1075         }
1076
1077         return NULL;
1078 }
1079
1080 static enum resp_states duplicate_request(struct rxe_qp *qp,
1081                                           struct rxe_pkt_info *pkt)
1082 {
1083         enum resp_states rc;
1084         u32 prev_psn = (qp->resp.ack_psn - 1) & BTH_PSN_MASK;
1085
1086         if (pkt->mask & RXE_SEND_MASK ||
1087             pkt->mask & RXE_WRITE_MASK) {
1088                 /* SEND. Ack again and cleanup. C9-105. */
1089                 if (bth_ack(pkt))
1090                         send_ack(qp, pkt, AETH_ACK_UNLIMITED, prev_psn);
1091                 rc = RESPST_CLEANUP;
1092                 goto out;
1093         } else if (pkt->mask & RXE_READ_MASK) {
1094                 struct resp_res *res;
1095
1096                 res = find_resource(qp, pkt->psn);
1097                 if (!res) {
1098                         /* Resource not found. Class D error.  Drop the
1099                          * request.
1100                          */
1101                         rc = RESPST_CLEANUP;
1102                         goto out;
1103                 } else {
1104                         /* Ensure this new request is the same as the previous
1105                          * one or a subset of it.
1106                          */
1107                         u64 iova = reth_va(pkt);
1108                         u32 resid = reth_len(pkt);
1109
1110                         if (iova < res->read.va_org ||
1111                             resid > res->read.length ||
1112                             (iova + resid) > (res->read.va_org +
1113                                               res->read.length)) {
1114                                 rc = RESPST_CLEANUP;
1115                                 goto out;
1116                         }
1117
1118                         if (reth_rkey(pkt) != res->read.rkey) {
1119                                 rc = RESPST_CLEANUP;
1120                                 goto out;
1121                         }
1122
1123                         res->cur_psn = pkt->psn;
1124                         res->state = (pkt->psn == res->first_psn) ?
1125                                         rdatm_res_state_new :
1126                                         rdatm_res_state_replay;
1127                         res->replay = 1;
1128
1129                         /* Reset the resource, except length. */
1130                         res->read.va_org = iova;
1131                         res->read.va = iova;
1132                         res->read.resid = resid;
1133
1134                         /* Replay the RDMA read reply. */
1135                         qp->resp.res = res;
1136                         rc = RESPST_READ_REPLY;
1137                         goto out;
1138                 }
1139         } else {
1140                 struct resp_res *res;
1141
1142                 /* Find the operation in our list of responder resources. */
1143                 res = find_resource(qp, pkt->psn);
1144                 if (res) {
1145                         skb_get(res->atomic.skb);
1146                         /* Resend the result. */
1147                         rc = rxe_xmit_packet(to_rdev(qp->ibqp.device), qp,
1148                                              pkt, res->atomic.skb);
1149                         if (rc) {
1150                                 pr_err("Failed resending result. This flow is not handled - skb ignored\n");
1151                                 rc = RESPST_CLEANUP;
1152                                 goto out;
1153                         }
1154                 }
1155
1156                 /* Resource not found. Class D error. Drop the request. */
1157                 rc = RESPST_CLEANUP;
1158                 goto out;
1159         }
1160 out:
1161         return rc;
1162 }
1163
1164 /* Process a class A or C. Both are treated the same in this implementation. */
1165 static void do_class_ac_error(struct rxe_qp *qp, u8 syndrome,
1166                               enum ib_wc_status status)
1167 {
1168         qp->resp.aeth_syndrome  = syndrome;
1169         qp->resp.status         = status;
1170
1171         /* indicate that we should go through the ERROR state */
1172         qp->resp.goto_error     = 1;
1173 }
1174
1175 static enum resp_states do_class_d1e_error(struct rxe_qp *qp)
1176 {
1177         /* UC */
1178         if (qp->srq) {
1179                 /* Class E */
1180                 qp->resp.drop_msg = 1;
1181                 if (qp->resp.wqe) {
1182                         qp->resp.status = IB_WC_REM_INV_REQ_ERR;
1183                         return RESPST_COMPLETE;
1184                 } else {
1185                         return RESPST_CLEANUP;
1186                 }
1187         } else {
1188                 /* Class D1. This packet may be the start of a
1189                  * new message and could be valid. The previous
1190                  * message is invalid and ignored. reset the
1191                  * recv wr to its original state
1192                  */
1193                 if (qp->resp.wqe) {
1194                         qp->resp.wqe->dma.resid = qp->resp.wqe->dma.length;
1195                         qp->resp.wqe->dma.cur_sge = 0;
1196                         qp->resp.wqe->dma.sge_offset = 0;
1197                         qp->resp.opcode = -1;
1198                 }
1199
1200                 if (qp->resp.mr) {
1201                         rxe_drop_ref(qp->resp.mr);
1202                         qp->resp.mr = NULL;
1203                 }
1204
1205                 return RESPST_CLEANUP;
1206         }
1207 }
1208
1209 static void rxe_drain_req_pkts(struct rxe_qp *qp, bool notify)
1210 {
1211         struct sk_buff *skb;
1212
1213         while ((skb = skb_dequeue(&qp->req_pkts))) {
1214                 rxe_drop_ref(qp);
1215                 kfree_skb(skb);
1216         }
1217
1218         if (notify)
1219                 return;
1220
1221         while (!qp->srq && qp->rq.queue && queue_head(qp->rq.queue))
1222                 advance_consumer(qp->rq.queue);
1223 }
1224
1225 int rxe_responder(void *arg)
1226 {
1227         struct rxe_qp *qp = (struct rxe_qp *)arg;
1228         struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
1229         enum resp_states state;
1230         struct rxe_pkt_info *pkt = NULL;
1231         int ret = 0;
1232
1233         rxe_add_ref(qp);
1234
1235         qp->resp.aeth_syndrome = AETH_ACK_UNLIMITED;
1236
1237         if (!qp->valid) {
1238                 ret = -EINVAL;
1239                 goto done;
1240         }
1241
1242         switch (qp->resp.state) {
1243         case QP_STATE_RESET:
1244                 state = RESPST_RESET;
1245                 break;
1246
1247         default:
1248                 state = RESPST_GET_REQ;
1249                 break;
1250         }
1251
1252         while (1) {
1253                 pr_debug("qp#%d state = %s\n", qp_num(qp),
1254                          resp_state_name[state]);
1255                 switch (state) {
1256                 case RESPST_GET_REQ:
1257                         state = get_req(qp, &pkt);
1258                         break;
1259                 case RESPST_CHK_PSN:
1260                         state = check_psn(qp, pkt);
1261                         break;
1262                 case RESPST_CHK_OP_SEQ:
1263                         state = check_op_seq(qp, pkt);
1264                         break;
1265                 case RESPST_CHK_OP_VALID:
1266                         state = check_op_valid(qp, pkt);
1267                         break;
1268                 case RESPST_CHK_RESOURCE:
1269                         state = check_resource(qp, pkt);
1270                         break;
1271                 case RESPST_CHK_LENGTH:
1272                         state = check_length(qp, pkt);
1273                         break;
1274                 case RESPST_CHK_RKEY:
1275                         state = check_rkey(qp, pkt);
1276                         break;
1277                 case RESPST_EXECUTE:
1278                         state = execute(qp, pkt);
1279                         break;
1280                 case RESPST_COMPLETE:
1281                         state = do_complete(qp, pkt);
1282                         break;
1283                 case RESPST_READ_REPLY:
1284                         state = read_reply(qp, pkt);
1285                         break;
1286                 case RESPST_ACKNOWLEDGE:
1287                         state = acknowledge(qp, pkt);
1288                         break;
1289                 case RESPST_CLEANUP:
1290                         state = cleanup(qp, pkt);
1291                         break;
1292                 case RESPST_DUPLICATE_REQUEST:
1293                         state = duplicate_request(qp, pkt);
1294                         break;
1295                 case RESPST_ERR_PSN_OUT_OF_SEQ:
1296                         /* RC only - Class B. Drop packet. */
1297                         send_ack(qp, pkt, AETH_NAK_PSN_SEQ_ERROR, qp->resp.psn);
1298                         state = RESPST_CLEANUP;
1299                         break;
1300
1301                 case RESPST_ERR_TOO_MANY_RDMA_ATM_REQ:
1302                 case RESPST_ERR_MISSING_OPCODE_FIRST:
1303                 case RESPST_ERR_MISSING_OPCODE_LAST_C:
1304                 case RESPST_ERR_UNSUPPORTED_OPCODE:
1305                 case RESPST_ERR_MISALIGNED_ATOMIC:
1306                         /* RC Only - Class C. */
1307                         do_class_ac_error(qp, AETH_NAK_INVALID_REQ,
1308                                           IB_WC_REM_INV_REQ_ERR);
1309                         state = RESPST_COMPLETE;
1310                         break;
1311
1312                 case RESPST_ERR_MISSING_OPCODE_LAST_D1E:
1313                         state = do_class_d1e_error(qp);
1314                         break;
1315                 case RESPST_ERR_RNR:
1316                         if (qp_type(qp) == IB_QPT_RC) {
1317                                 rxe_counter_inc(rxe, RXE_CNT_SND_RNR);
1318                                 /* RC - class B */
1319                                 send_ack(qp, pkt, AETH_RNR_NAK |
1320                                          (~AETH_TYPE_MASK &
1321                                          qp->attr.min_rnr_timer),
1322                                          pkt->psn);
1323                         } else {
1324                                 /* UD/UC - class D */
1325                                 qp->resp.drop_msg = 1;
1326                         }
1327                         state = RESPST_CLEANUP;
1328                         break;
1329
1330                 case RESPST_ERR_RKEY_VIOLATION:
1331                         if (qp_type(qp) == IB_QPT_RC) {
1332                                 /* Class C */
1333                                 do_class_ac_error(qp, AETH_NAK_REM_ACC_ERR,
1334                                                   IB_WC_REM_ACCESS_ERR);
1335                                 state = RESPST_COMPLETE;
1336                         } else {
1337                                 qp->resp.drop_msg = 1;
1338                                 if (qp->srq) {
1339                                         /* UC/SRQ Class D */
1340                                         qp->resp.status = IB_WC_REM_ACCESS_ERR;
1341                                         state = RESPST_COMPLETE;
1342                                 } else {
1343                                         /* UC/non-SRQ Class E. */
1344                                         state = RESPST_CLEANUP;
1345                                 }
1346                         }
1347                         break;
1348
1349                 case RESPST_ERR_LENGTH:
1350                         if (qp_type(qp) == IB_QPT_RC) {
1351                                 /* Class C */
1352                                 do_class_ac_error(qp, AETH_NAK_INVALID_REQ,
1353                                                   IB_WC_REM_INV_REQ_ERR);
1354                                 state = RESPST_COMPLETE;
1355                         } else if (qp->srq) {
1356                                 /* UC/UD - class E */
1357                                 qp->resp.status = IB_WC_REM_INV_REQ_ERR;
1358                                 state = RESPST_COMPLETE;
1359                         } else {
1360                                 /* UC/UD - class D */
1361                                 qp->resp.drop_msg = 1;
1362                                 state = RESPST_CLEANUP;
1363                         }
1364                         break;
1365
1366                 case RESPST_ERR_MALFORMED_WQE:
1367                         /* All, Class A. */
1368                         do_class_ac_error(qp, AETH_NAK_REM_OP_ERR,
1369                                           IB_WC_LOC_QP_OP_ERR);
1370                         state = RESPST_COMPLETE;
1371                         break;
1372
1373                 case RESPST_ERR_CQ_OVERFLOW:
1374                         /* All - Class G */
1375                         state = RESPST_ERROR;
1376                         break;
1377
1378                 case RESPST_DONE:
1379                         if (qp->resp.goto_error) {
1380                                 state = RESPST_ERROR;
1381                                 break;
1382                         }
1383
1384                         goto done;
1385
1386                 case RESPST_EXIT:
1387                         if (qp->resp.goto_error) {
1388                                 state = RESPST_ERROR;
1389                                 break;
1390                         }
1391
1392                         goto exit;
1393
1394                 case RESPST_RESET:
1395                         rxe_drain_req_pkts(qp, false);
1396                         qp->resp.wqe = NULL;
1397                         goto exit;
1398
1399                 case RESPST_ERROR:
1400                         qp->resp.goto_error = 0;
1401                         pr_warn("qp#%d moved to error state\n", qp_num(qp));
1402                         rxe_qp_error(qp);
1403                         goto exit;
1404
1405                 default:
1406                         WARN_ON_ONCE(1);
1407                 }
1408         }
1409
1410 exit:
1411         ret = -EAGAIN;
1412 done:
1413         rxe_drop_ref(qp);
1414         return ret;
1415 }