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
 153         switch (qp_type(qp)) {
 154         case IB_QPT_RC:
 155                 if (diff > 0) {
 156                         if (qp->resp.sent_psn_nak)
 157                                 return RESPST_CLEANUP;
 158
 159                         qp->resp.sent_psn_nak = 1;
 160                         return RESPST_ERR_PSN_OUT_OF_SEQ;
 161
 162                 } else if (diff < 0) {
 163                         return RESPST_DUPLICATE_REQUEST;
 164                 }
 165
 166                 if (qp->resp.sent_psn_nak)
 167                         qp->resp.sent_psn_nak = 0;
 168
 169                 break;
 170
 171         case IB_QPT_UC:
 172                 if (qp->resp.drop_msg || diff != 0) {
 173                         if (pkt->mask & RXE_START_MASK) {
 174                                 qp->resp.drop_msg = 0;
 175                                 return RESPST_CHK_OP_SEQ;
 176                         }
 177
 178                         qp->resp.drop_msg = 1;
 179                         return RESPST_CLEANUP;
 180                 }
 181                 break;
 182         default:
 183                 break;
 184         }
 185
 186         return RESPST_CHK_OP_SEQ;
 187 }
 188
 189 static enum resp_states check_op_seq(struct rxe_qp *qp,
 190                                      struct rxe_pkt_info *pkt)
 191 {
 192         switch (qp_type(qp)) {
 193         case IB_QPT_RC:
 194                 switch (qp->resp.opcode) {
 195                 case IB_OPCODE_RC_SEND_FIRST:
 196                 case IB_OPCODE_RC_SEND_MIDDLE:
 197                         switch (pkt->opcode) {
 198                         case IB_OPCODE_RC_SEND_MIDDLE:
 199                         case IB_OPCODE_RC_SEND_LAST:
 200                         case IB_OPCODE_RC_SEND_LAST_WITH_IMMEDIATE:
 201                         case IB_OPCODE_RC_SEND_LAST_WITH_INVALIDATE:
 202                                 return RESPST_CHK_OP_VALID;
 203                         default:
 204                                 return RESPST_ERR_MISSING_OPCODE_LAST_C;
 205                         }
 206
 207                 case IB_OPCODE_RC_RDMA_WRITE_FIRST:
 208                 case IB_OPCODE_RC_RDMA_WRITE_MIDDLE:
 209                         switch (pkt->opcode) {
 210                         case IB_OPCODE_RC_RDMA_WRITE_MIDDLE:
 211                         case IB_OPCODE_RC_RDMA_WRITE_LAST:
 212                         case IB_OPCODE_RC_RDMA_WRITE_LAST_WITH_IMMEDIATE:
 213                                 return RESPST_CHK_OP_VALID;
 214                         default:
 215                                 return RESPST_ERR_MISSING_OPCODE_LAST_C;
 216                         }
 217
 218                 default:
 219                         switch (pkt->opcode) {
 220                         case IB_OPCODE_RC_SEND_MIDDLE:
 221                         case IB_OPCODE_RC_SEND_LAST:
 222                         case IB_OPCODE_RC_SEND_LAST_WITH_IMMEDIATE:
 223                         case IB_OPCODE_RC_SEND_LAST_WITH_INVALIDATE:
 224                         case IB_OPCODE_RC_RDMA_WRITE_MIDDLE:
 225                         case IB_OPCODE_RC_RDMA_WRITE_LAST:
 226                         case IB_OPCODE_RC_RDMA_WRITE_LAST_WITH_IMMEDIATE:
 227                                 return RESPST_ERR_MISSING_OPCODE_FIRST;
 228                         default:
 229                                 return RESPST_CHK_OP_VALID;
 230                         }
 231                 }
 232                 break;
 233
 234         case IB_QPT_UC:
 235                 switch (qp->resp.opcode) {
 236                 case IB_OPCODE_UC_SEND_FIRST:
 237                 case IB_OPCODE_UC_SEND_MIDDLE:
 238                         switch (pkt->opcode) {
 239                         case IB_OPCODE_UC_SEND_MIDDLE:
 240                         case IB_OPCODE_UC_SEND_LAST:
 241                         case IB_OPCODE_UC_SEND_LAST_WITH_IMMEDIATE:
 242                                 return RESPST_CHK_OP_VALID;
 243                         default:
 244                                 return RESPST_ERR_MISSING_OPCODE_LAST_D1E;
 245                         }
 246
 247                 case IB_OPCODE_UC_RDMA_WRITE_FIRST:
 248                 case IB_OPCODE_UC_RDMA_WRITE_MIDDLE:
 249                         switch (pkt->opcode) {
 250                         case IB_OPCODE_UC_RDMA_WRITE_MIDDLE:
 251                         case IB_OPCODE_UC_RDMA_WRITE_LAST:
 252                         case IB_OPCODE_UC_RDMA_WRITE_LAST_WITH_IMMEDIATE:
 253                                 return RESPST_CHK_OP_VALID;
 254                         default:
 255                                 return RESPST_ERR_MISSING_OPCODE_LAST_D1E;
 256                         }
 257
 258                 default:
 259                         switch (pkt->opcode) {
 260                         case IB_OPCODE_UC_SEND_MIDDLE:
 261                         case IB_OPCODE_UC_SEND_LAST:
 262                         case IB_OPCODE_UC_SEND_LAST_WITH_IMMEDIATE:
 263                         case IB_OPCODE_UC_RDMA_WRITE_MIDDLE:
 264                         case IB_OPCODE_UC_RDMA_WRITE_LAST:
 265                         case IB_OPCODE_UC_RDMA_WRITE_LAST_WITH_IMMEDIATE:
 266                                 qp->resp.drop_msg = 1;
 267                                 return RESPST_CLEANUP;
 268                         default:
 269                                 return RESPST_CHK_OP_VALID;
 270                         }
 271                 }
 272                 break;
 273
 274         default:
 275                 return RESPST_CHK_OP_VALID;
 276         }
 277 }
 278
 279 static enum resp_states check_op_valid(struct rxe_qp *qp,
 280                                        struct rxe_pkt_info *pkt)
 281 {
 282         switch (qp_type(qp)) {
 283         case IB_QPT_RC:
 284                 if (((pkt->mask & RXE_READ_MASK) &&
 285                      !(qp->attr.qp_access_flags & IB_ACCESS_REMOTE_READ)) ||
 286                     ((pkt->mask & RXE_WRITE_MASK) &&
 287                      !(qp->attr.qp_access_flags & IB_ACCESS_REMOTE_WRITE)) ||
 288                     ((pkt->mask & RXE_ATOMIC_MASK) &&
 289                      !(qp->attr.qp_access_flags & IB_ACCESS_REMOTE_ATOMIC))) {
 290                         return RESPST_ERR_UNSUPPORTED_OPCODE;
 291                 }
 292
 293                 break;
 294
 295         case IB_QPT_UC:
 296                 if ((pkt->mask & RXE_WRITE_MASK) &&
 297                     !(qp->attr.qp_access_flags & IB_ACCESS_REMOTE_WRITE)) {
 298                         qp->resp.drop_msg = 1;
 299                         return RESPST_CLEANUP;
 300                 }
 301
 302                 break;
 303
 304         case IB_QPT_UD:
 305         case IB_QPT_SMI:
 306         case IB_QPT_GSI:
 307                 break;
 308
 309         default:
 310                 WARN_ON(1);
 311                 break;
 312         }
 313
 314         return RESPST_CHK_RESOURCE;
 315 }
 316
 317 static enum resp_states get_srq_wqe(struct rxe_qp *qp)
 318 {
 319         struct rxe_srq *srq = qp->srq;
 320         struct rxe_queue *q = srq->rq.queue;
 321         struct rxe_recv_wqe *wqe;
 322         struct ib_event ev;
 323
 324         if (srq->error)
 325                 return RESPST_ERR_RNR;
 326
 327         spin_lock_bh(&srq->rq.consumer_lock);
 328
 329         wqe = queue_head(q);
 330         if (!wqe) {
 331                 spin_unlock_bh(&srq->rq.consumer_lock);
 332                 return RESPST_ERR_RNR;
 333         }
 334
 335         /* note kernel and user space recv wqes have same size */
 336         memcpy(&qp->resp.srq_wqe, wqe, sizeof(qp->resp.srq_wqe));
 337
 338         qp->resp.wqe = &qp->resp.srq_wqe.wqe;
 339         advance_consumer(q);
 340
 341         if (srq->limit && srq->ibsrq.event_handler &&
 342             (queue_count(q) < srq->limit)) {
 343                 srq->limit = 0;
 344                 goto event;
 345         }
 346
 347         spin_unlock_bh(&srq->rq.consumer_lock);
 348         return RESPST_CHK_LENGTH;
 349
 350 event:
 351         spin_unlock_bh(&srq->rq.consumer_lock);
 352         ev.device = qp->ibqp.device;
 353         ev.element.srq = qp->ibqp.srq;
 354         ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
 355         srq->ibsrq.event_handler(&ev, srq->ibsrq.srq_context);
 356         return RESPST_CHK_LENGTH;
 357 }
 358
 359 static enum resp_states check_resource(struct rxe_qp *qp,
 360                                        struct rxe_pkt_info *pkt)
 361 {
 362         struct rxe_srq *srq = qp->srq;
 363
 364         if (qp->resp.state == QP_STATE_ERROR) {
 365                 if (qp->resp.wqe) {
 366                         qp->resp.status = IB_WC_WR_FLUSH_ERR;
 367                         return RESPST_COMPLETE;
 368                 } else if (!srq) {
 369                         qp->resp.wqe = queue_head(qp->rq.queue);
 370                         if (qp->resp.wqe) {
 371                                 qp->resp.status = IB_WC_WR_FLUSH_ERR;
 372                                 return RESPST_COMPLETE;
 373                         } else {
 374                                 return RESPST_EXIT;
 375                         }
 376                 } else {
 377                         return RESPST_EXIT;
 378                 }
 379         }
 380
 381         if (pkt->mask & RXE_READ_OR_ATOMIC) {
 382                 /* it is the requesters job to not send
 383                  * too many read/atomic ops, we just
 384                  * recycle the responder resource queue
 385                  */
 386                 if (likely(qp->attr.max_dest_rd_atomic > 0))
 387                         return RESPST_CHK_LENGTH;
 388                 else
 389                         return RESPST_ERR_TOO_MANY_RDMA_ATM_REQ;
 390         }
 391
 392         if (pkt->mask & RXE_RWR_MASK) {
 393                 if (srq)
 394                         return get_srq_wqe(qp);
 395
 396                 qp->resp.wqe = queue_head(qp->rq.queue);
 397                 return (qp->resp.wqe) ? RESPST_CHK_LENGTH : RESPST_ERR_RNR;
 398         }
 399
 400         return RESPST_CHK_LENGTH;
 401 }
 402
 403 static enum resp_states check_length(struct rxe_qp *qp,
 404                                      struct rxe_pkt_info *pkt)
 405 {
 406         switch (qp_type(qp)) {
 407         case IB_QPT_RC:
 408                 return RESPST_CHK_RKEY;
 409
 410         case IB_QPT_UC:
 411                 return RESPST_CHK_RKEY;
 412
 413         default:
 414                 return RESPST_CHK_RKEY;
 415         }
 416 }
 417
 418 static enum resp_states check_rkey(struct rxe_qp *qp,
 419                                    struct rxe_pkt_info *pkt)
 420 {
 421         struct rxe_mem *mem;
 422         u64 va;
 423         u32 rkey;
 424         u32 resid;
 425         u32 pktlen;
 426         int mtu = qp->mtu;
 427         enum resp_states state;
 428         int access;
 429
 430         if (pkt->mask & (RXE_READ_MASK | RXE_WRITE_MASK)) {
 431                 if (pkt->mask & RXE_RETH_MASK) {
 432                         qp->resp.va = reth_va(pkt);
 433                         qp->resp.rkey = reth_rkey(pkt);
 434                         qp->resp.resid = reth_len(pkt);
 435                 }
 436                 access = (pkt->mask & RXE_READ_MASK) ? IB_ACCESS_REMOTE_READ
 437                                                      : IB_ACCESS_REMOTE_WRITE;
 438         } else if (pkt->mask & RXE_ATOMIC_MASK) {
 439                 qp->resp.va = atmeth_va(pkt);
 440                 qp->resp.rkey = atmeth_rkey(pkt);
 441                 qp->resp.resid = sizeof(u64);
 442                 access = IB_ACCESS_REMOTE_ATOMIC;
 443         } else {
 444                 return RESPST_EXECUTE;
 445         }
 446
 447         va      = qp->resp.va;
 448         rkey    = qp->resp.rkey;
 449         resid   = qp->resp.resid;
 450         pktlen  = payload_size(pkt);
 451
 452         mem = lookup_mem(qp->pd, access, rkey, lookup_remote);
 453         if (!mem) {
 454                 state = RESPST_ERR_RKEY_VIOLATION;
 455                 goto err1;
 456         }
 457
 458         if (unlikely(mem->state == RXE_MEM_STATE_FREE)) {
 459                 state = RESPST_ERR_RKEY_VIOLATION;
 460                 goto err1;
 461         }
 462
 463         if (mem_check_range(mem, va, resid)) {
 464                 state = RESPST_ERR_RKEY_VIOLATION;
 465                 goto err2;
 466         }
 467
 468         if (pkt->mask & RXE_WRITE_MASK)  {
 469                 if (resid > mtu) {
 470                         if (pktlen != mtu || bth_pad(pkt)) {
 471                                 state = RESPST_ERR_LENGTH;
 472                                 goto err2;
 473                         }
 474
 475                         resid = mtu;
 476                 } else {
 477                         if (pktlen != resid) {
 478                                 state = RESPST_ERR_LENGTH;
 479                                 goto err2;
 480                         }
 481                         if ((bth_pad(pkt) != (0x3 & (-resid)))) {
 482                                 /* This case may not be exactly that
 483                                  * but nothing else fits.
 484                                  */
 485                                 state = RESPST_ERR_LENGTH;
 486                                 goto err2;
 487                         }
 488                 }
 489         }
 490
 491         WARN_ON(qp->resp.mr);
 492
 493         qp->resp.mr = mem;
 494         return RESPST_EXECUTE;
 495
 496 err2:
 497         rxe_drop_ref(mem);
 498 err1:
 499         return state;
 500 }
 501
 502 static enum resp_states send_data_in(struct rxe_qp *qp, void *data_addr,
 503                                      int data_len)
 504 {
 505         int err;
 506         struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
 507
 508         err = copy_data(rxe, qp->pd, IB_ACCESS_LOCAL_WRITE, &qp->resp.wqe->dma,
 509                         data_addr, data_len, to_mem_obj, NULL);
 510         if (unlikely(err))
 511                 return (err == -ENOSPC) ? RESPST_ERR_LENGTH
 512                                         : RESPST_ERR_MALFORMED_WQE;
 513
 514         return RESPST_NONE;
 515 }
 516
 517 static enum resp_states write_data_in(struct rxe_qp *qp,
 518                                       struct rxe_pkt_info *pkt)
 519 {
 520         enum resp_states rc = RESPST_NONE;
 521         int     err;
 522         int data_len = payload_size(pkt);
 523
 524         err = rxe_mem_copy(qp->resp.mr, qp->resp.va, payload_addr(pkt),
 525                            data_len, to_mem_obj, NULL);
 526         if (err) {
 527                 rc = RESPST_ERR_RKEY_VIOLATION;
 528                 goto out;
 529         }
 530
 531         qp->resp.va += data_len;
 532         qp->resp.resid -= data_len;
 533
 534 out:
 535         return rc;
 536 }
 537
 538 /* Guarantee atomicity of atomic operations at the machine level. */
 539 static DEFINE_SPINLOCK(atomic_ops_lock);
 540
 541 static enum resp_states process_atomic(struct rxe_qp *qp,
 542                                        struct rxe_pkt_info *pkt)
 543 {
 544         u64 iova = atmeth_va(pkt);
 545         u64 *vaddr;
 546         enum resp_states ret;
 547         struct rxe_mem *mr = qp->resp.mr;
 548
 549         if (mr->state != RXE_MEM_STATE_VALID) {
 550                 ret = RESPST_ERR_RKEY_VIOLATION;
 551                 goto out;
 552         }
 553
 554         vaddr = iova_to_vaddr(mr, iova, sizeof(u64));
 555
 556         /* check vaddr is 8 bytes aligned. */
 557         if (!vaddr || (uintptr_t)vaddr & 7) {
 558                 ret = RESPST_ERR_MISALIGNED_ATOMIC;
 559                 goto out;
 560         }
 561
 562         spin_lock_bh(&atomic_ops_lock);
 563
 564         qp->resp.atomic_orig = *vaddr;
 565
 566         if (pkt->opcode == IB_OPCODE_RC_COMPARE_SWAP ||
 567             pkt->opcode == IB_OPCODE_RD_COMPARE_SWAP) {
 568                 if (*vaddr == atmeth_comp(pkt))
 569                         *vaddr = atmeth_swap_add(pkt);
 570         } else {
 571                 *vaddr += atmeth_swap_add(pkt);
 572         }
 573
 574         spin_unlock_bh(&atomic_ops_lock);
 575
 576         ret = RESPST_NONE;
 577 out:
 578         return ret;
 579 }
 580
 581 static struct sk_buff *prepare_ack_packet(struct rxe_qp *qp,
 582                                           struct rxe_pkt_info *pkt,
 583                                           struct rxe_pkt_info *ack,
 584                                           int opcode,
 585                                           int payload,
 586                                           u32 psn,
 587                                           u8 syndrome,
 588                                           u32 *crcp)
 589 {
 590         struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
 591         struct sk_buff *skb;
 592         u32 crc = 0;
 593         u32 *p;
 594         int paylen;
 595         int pad;
 596         int err;
 597
 598         /*
 599          * allocate packet
 600          */
 601         pad = (-payload) & 0x3;
 602         paylen = rxe_opcode[opcode].length + payload + pad + RXE_ICRC_SIZE;
 603
 604         skb = rxe->ifc_ops->init_packet(rxe, &qp->pri_av, paylen, ack);
 605         if (!skb)
 606                 return NULL;
 607
 608         ack->qp = qp;
 609         ack->opcode = opcode;
 610         ack->mask = rxe_opcode[opcode].mask;
 611         ack->offset = pkt->offset;
 612         ack->paylen = paylen;
 613
 614         /* fill in bth using the request packet headers */
 615         memcpy(ack->hdr, pkt->hdr, pkt->offset + RXE_BTH_BYTES);
 616
 617         bth_set_opcode(ack, opcode);
 618         bth_set_qpn(ack, qp->attr.dest_qp_num);
 619         bth_set_pad(ack, pad);
 620         bth_set_se(ack, 0);
 621         bth_set_psn(ack, psn);
 622         bth_set_ack(ack, 0);
 623         ack->psn = psn;
 624
 625         if (ack->mask & RXE_AETH_MASK) {
 626                 aeth_set_syn(ack, syndrome);
 627                 aeth_set_msn(ack, qp->resp.msn);
 628         }
 629
 630         if (ack->mask & RXE_ATMACK_MASK)
 631                 atmack_set_orig(ack, qp->resp.atomic_orig);
 632
 633         err = rxe->ifc_ops->prepare(rxe, ack, skb, &crc);
 634         if (err) {
 635                 kfree_skb(skb);
 636                 return NULL;
 637         }
 638
 639         if (crcp) {
 640                 /* CRC computation will be continued by the caller */
 641                 *crcp = crc;
 642         } else {
 643                 p = payload_addr(ack) + payload + bth_pad(ack);
 644                 *p = ~crc;
 645         }
 646
 647         return skb;
 648 }
 649
 650 /* RDMA read response. If res is not NULL, then we have a current RDMA request
 651  * being processed or replayed.
 652  */
 653 static enum resp_states read_reply(struct rxe_qp *qp,
 654                                    struct rxe_pkt_info *req_pkt)
 655 {
 656         struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
 657         struct rxe_pkt_info ack_pkt;
 658         struct sk_buff *skb;
 659         int mtu = qp->mtu;
 660         enum resp_states state;
 661         int payload;
 662         int opcode;
 663         int err;
 664         struct resp_res *res = qp->resp.res;
 665         u32 icrc;
 666         u32 *p;
 667
 668         if (!res) {
 669                 /* This is the first time we process that request. Get a
 670                  * resource
 671                  */
 672                 res = &qp->resp.resources[qp->resp.res_head];
 673
 674                 free_rd_atomic_resource(qp, res);
 675                 rxe_advance_resp_resource(qp);
 676
 677                 res->type               = RXE_READ_MASK;
 678
 679                 res->read.va            = qp->resp.va;
 680                 res->read.va_org        = qp->resp.va;
 681
 682                 res->first_psn          = req_pkt->psn;
 683                 res->last_psn           = req_pkt->psn +
 684                                           (reth_len(req_pkt) + mtu - 1) /
 685                                           mtu - 1;
 686                 res->cur_psn            = req_pkt->psn;
 687
 688                 res->read.resid         = qp->resp.resid;
 689                 res->read.length        = qp->resp.resid;
 690                 res->read.rkey          = qp->resp.rkey;
 691
 692                 /* note res inherits the reference to mr from qp */
 693                 res->read.mr            = qp->resp.mr;
 694                 qp->resp.mr             = NULL;
 695
 696                 qp->resp.res            = res;
 697                 res->state              = rdatm_res_state_new;
 698         }
 699
 700         if (res->state == rdatm_res_state_new) {
 701                 if (res->read.resid <= mtu)
 702                         opcode = IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY;
 703                 else
 704                         opcode = IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST;
 705         } else {
 706                 if (res->read.resid > mtu)
 707                         opcode = IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE;
 708                 else
 709                         opcode = IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST;
 710         }
 711
 712         res->state = rdatm_res_state_next;
 713
 714         payload = min_t(int, res->read.resid, mtu);
 715
 716         skb = prepare_ack_packet(qp, req_pkt, &ack_pkt, opcode, payload,
 717                                  res->cur_psn, AETH_ACK_UNLIMITED, &icrc);
 718         if (!skb)
 719                 return RESPST_ERR_RNR;
 720
 721         err = rxe_mem_copy(res->read.mr, res->read.va, payload_addr(&ack_pkt),
 722                            payload, from_mem_obj, &icrc);
 723         if (err)
 724                 pr_err("Failed copying memory\n");
 725
 726         p = payload_addr(&ack_pkt) + payload + bth_pad(&ack_pkt);
 727         *p = ~icrc;
 728
 729         err = rxe_xmit_packet(rxe, qp, &ack_pkt, skb);
 730         if (err) {
 731                 pr_err("Failed sending RDMA reply.\n");
 732                 kfree_skb(skb);
 733                 return RESPST_ERR_RNR;
 734         }
 735
 736         res->read.va += payload;
 737         res->read.resid -= payload;
 738         res->cur_psn = (res->cur_psn + 1) & BTH_PSN_MASK;
 739
 740         if (res->read.resid > 0) {
 741                 state = RESPST_DONE;
 742         } else {
 743                 qp->resp.res = NULL;
 744                 qp->resp.opcode = -1;
 745                 qp->resp.psn = res->cur_psn;
 746                 state = RESPST_CLEANUP;
 747         }
 748
 749         return state;
 750 }
 751
 752 static void build_rdma_network_hdr(union rdma_network_hdr *hdr,
 753                                    struct rxe_pkt_info *pkt)
 754 {
 755         struct sk_buff *skb = PKT_TO_SKB(pkt);
 756
 757         memset(hdr, 0, sizeof(*hdr));
 758         if (skb->protocol == htons(ETH_P_IP))
 759                 memcpy(&hdr->roce4grh, ip_hdr(skb), sizeof(hdr->roce4grh));
 760         else if (skb->protocol == htons(ETH_P_IPV6))
 761                 memcpy(&hdr->ibgrh, ipv6_hdr(skb), sizeof(hdr->ibgrh));
 762 }
 763
 764 /* Executes a new request. A retried request never reach that function (send
 765  * and writes are discarded, and reads and atomics are retried elsewhere.
 766  */
 767 static enum resp_states execute(struct rxe_qp *qp, struct rxe_pkt_info *pkt)
 768 {
 769         enum resp_states err;
 770
 771         if (pkt->mask & RXE_SEND_MASK) {
 772                 if (qp_type(qp) == IB_QPT_UD ||
 773                     qp_type(qp) == IB_QPT_SMI ||
 774                     qp_type(qp) == IB_QPT_GSI) {
 775                         union rdma_network_hdr hdr;
 776
 777                         build_rdma_network_hdr(&hdr, pkt);
 778
 779                         err = send_data_in(qp, &hdr, sizeof(hdr));
 780                         if (err)
 781                                 return err;
 782                 }
 783                 err = send_data_in(qp, payload_addr(pkt), payload_size(pkt));
 784                 if (err)
 785                         return err;
 786         } else if (pkt->mask & RXE_WRITE_MASK) {
 787                 err = write_data_in(qp, pkt);
 788                 if (err)
 789                         return err;
 790         } else if (pkt->mask & RXE_READ_MASK) {
 791                 /* For RDMA Read we can increment the msn now. See C9-148. */
 792                 qp->resp.msn++;
 793                 return RESPST_READ_REPLY;
 794         } else if (pkt->mask & RXE_ATOMIC_MASK) {
 795                 err = process_atomic(qp, pkt);
 796                 if (err)
 797                         return err;
 798         } else
 799                 /* Unreachable */
 800                 WARN_ON(1);
 801
 802         /* We successfully processed this new request. */
 803         qp->resp.msn++;
 804
 805         /* next expected psn, read handles this separately */
 806         qp->resp.psn = (pkt->psn + 1) & BTH_PSN_MASK;
 807
 808         qp->resp.opcode = pkt->opcode;
 809         qp->resp.status = IB_WC_SUCCESS;
 810
 811         if (pkt->mask & RXE_COMP_MASK)
 812                 return RESPST_COMPLETE;
 813         else if (qp_type(qp) == IB_QPT_RC)
 814                 return RESPST_ACKNOWLEDGE;
 815         else
 816                 return RESPST_CLEANUP;
 817 }
 818
 819 static enum resp_states do_complete(struct rxe_qp *qp,
 820                                     struct rxe_pkt_info *pkt)
 821 {
 822         struct rxe_cqe cqe;
 823         struct ib_wc *wc = &cqe.ibwc;
 824         struct ib_uverbs_wc *uwc = &cqe.uibwc;
 825         struct rxe_recv_wqe *wqe = qp->resp.wqe;
 826
 827         if (unlikely(!wqe))
 828                 return RESPST_CLEANUP;
 829
 830         memset(&cqe, 0, sizeof(cqe));
 831
 832         wc->wr_id               = wqe->wr_id;
 833         wc->status              = qp->resp.status;
 834         wc->qp                  = &qp->ibqp;
 835
 836         /* fields after status are not required for errors */
 837         if (wc->status == IB_WC_SUCCESS) {
 838                 wc->opcode = (pkt->mask & RXE_IMMDT_MASK &&
 839                                 pkt->mask & RXE_WRITE_MASK) ?
 840                                         IB_WC_RECV_RDMA_WITH_IMM : IB_WC_RECV;
 841                 wc->vendor_err = 0;
 842                 wc->byte_len = wqe->dma.length - wqe->dma.resid;
 843
 844                 /* fields after byte_len are different between kernel and user
 845                  * space
 846                  */
 847                 if (qp->rcq->is_user) {
 848                         uwc->wc_flags = IB_WC_GRH;
 849
 850                         if (pkt->mask & RXE_IMMDT_MASK) {
 851                                 uwc->wc_flags |= IB_WC_WITH_IMM;
 852                                 uwc->ex.imm_data =
 853                                         (__u32 __force)immdt_imm(pkt);
 854                         }
 855
 856                         if (pkt->mask & RXE_IETH_MASK) {
 857                                 uwc->wc_flags |= IB_WC_WITH_INVALIDATE;
 858                                 uwc->ex.invalidate_rkey = ieth_rkey(pkt);
 859                         }
 860
 861                         uwc->qp_num             = qp->ibqp.qp_num;
 862
 863                         if (pkt->mask & RXE_DETH_MASK)
 864                                 uwc->src_qp = deth_sqp(pkt);
 865
 866                         uwc->port_num           = qp->attr.port_num;
 867                 } else {
 868                         struct sk_buff *skb = PKT_TO_SKB(pkt);
 869
 870                         wc->wc_flags = IB_WC_GRH | IB_WC_WITH_NETWORK_HDR_TYPE;
 871                         if (skb->protocol == htons(ETH_P_IP))
 872                                 wc->network_hdr_type = RDMA_NETWORK_IPV4;
 873                         else
 874                                 wc->network_hdr_type = RDMA_NETWORK_IPV6;
 875
 876                         if (pkt->mask & RXE_IMMDT_MASK) {
 877                                 wc->wc_flags |= IB_WC_WITH_IMM;
 878                                 wc->ex.imm_data = immdt_imm(pkt);
 879                         }
 880
 881                         if (pkt->mask & RXE_IETH_MASK) {
 882                                 struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
 883                                 struct rxe_mem *rmr;
 884
 885                                 wc->wc_flags |= IB_WC_WITH_INVALIDATE;
 886                                 wc->ex.invalidate_rkey = ieth_rkey(pkt);
 887
 888                                 rmr = rxe_pool_get_index(&rxe->mr_pool,
 889                                                          wc->ex.invalidate_rkey >> 8);
 890                                 if (unlikely(!rmr)) {
 891                                         pr_err("Bad rkey %#x invalidation\n",
 892                                                wc->ex.invalidate_rkey);
 893                                         return RESPST_ERROR;
 894                                 }
 895                                 rmr->state = RXE_MEM_STATE_FREE;
 896                         }
 897
 898                         wc->qp                  = &qp->ibqp;
 899
 900                         if (pkt->mask & RXE_DETH_MASK)
 901                                 wc->src_qp = deth_sqp(pkt);
 902
 903                         wc->port_num            = qp->attr.port_num;
 904                 }
 905         }
 906
 907         /* have copy for srq and reference for !srq */
 908         if (!qp->srq)
 909                 advance_consumer(qp->rq.queue);
 910
 911         qp->resp.wqe = NULL;
 912
 913         if (rxe_cq_post(qp->rcq, &cqe, pkt ? bth_se(pkt) : 1))
 914                 return RESPST_ERR_CQ_OVERFLOW;
 915
 916         if (qp->resp.state == QP_STATE_ERROR)
 917                 return RESPST_CHK_RESOURCE;
 918
 919         if (!pkt)
 920                 return RESPST_DONE;
 921         else if (qp_type(qp) == IB_QPT_RC)
 922                 return RESPST_ACKNOWLEDGE;
 923         else
 924                 return RESPST_CLEANUP;
 925 }
 926
 927 static int send_ack(struct rxe_qp *qp, struct rxe_pkt_info *pkt,
 928                     u8 syndrome, u32 psn)
 929 {
 930         int err = 0;
 931         struct rxe_pkt_info ack_pkt;
 932         struct sk_buff *skb;
 933         struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
 934
 935         skb = prepare_ack_packet(qp, pkt, &ack_pkt, IB_OPCODE_RC_ACKNOWLEDGE,
 936                                  0, psn, syndrome, NULL);
 937         if (!skb) {
 938                 err = -ENOMEM;
 939                 goto err1;
 940         }
 941
 942         err = rxe_xmit_packet(rxe, qp, &ack_pkt, skb);
 943         if (err) {
 944                 pr_err_ratelimited("Failed sending ack\n");
 945                 kfree_skb(skb);
 946         }
 947
 948 err1:
 949         return err;
 950 }
 951
 952 static int send_atomic_ack(struct rxe_qp *qp, struct rxe_pkt_info *pkt,
 953                            u8 syndrome)
 954 {
 955         int rc = 0;
 956         struct rxe_pkt_info ack_pkt;
 957         struct sk_buff *skb;
 958         struct sk_buff *skb_copy;
 959         struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
 960         struct resp_res *res;
 961
 962         skb = prepare_ack_packet(qp, pkt, &ack_pkt,
 963                                  IB_OPCODE_RC_ATOMIC_ACKNOWLEDGE, 0, pkt->psn,
 964                                  syndrome, NULL);
 965         if (!skb) {
 966                 rc = -ENOMEM;
 967                 goto out;
 968         }
 969
 970         skb_copy = skb_clone(skb, GFP_ATOMIC);
 971         if (skb_copy)
 972                 rxe_add_ref(qp); /* for the new SKB */
 973         else {
 974                 pr_warn("Could not clone atomic response\n");
 975                 rc = -ENOMEM;
 976                 goto out;
 977         }
 978
 979         res = &qp->resp.resources[qp->resp.res_head];
 980         free_rd_atomic_resource(qp, res);
 981         rxe_advance_resp_resource(qp);
 982
 983         memcpy(SKB_TO_PKT(skb), &ack_pkt, sizeof(skb->cb));
 984
 985         res->type = RXE_ATOMIC_MASK;
 986         res->atomic.skb = skb;
 987         res->first_psn = ack_pkt.psn;
 988         res->last_psn  = ack_pkt.psn;
 989         res->cur_psn   = ack_pkt.psn;
 990
 991         rc = rxe_xmit_packet(rxe, qp, &ack_pkt, skb_copy);
 992         if (rc) {
 993                 pr_err_ratelimited("Failed sending ack\n");
 994                 rxe_drop_ref(qp);
 995                 kfree_skb(skb_copy);
 996         }
 997
 998 out:
 999         return rc;
1000 }
1001
1002 static enum resp_states acknowledge(struct rxe_qp *qp,
1003                                     struct rxe_pkt_info *pkt)
1004 {
1005         if (qp_type(qp) != IB_QPT_RC)
1006                 return RESPST_CLEANUP;
1007
1008         if (qp->resp.aeth_syndrome != AETH_ACK_UNLIMITED)
1009                 send_ack(qp, pkt, qp->resp.aeth_syndrome, pkt->psn);
1010         else if (pkt->mask & RXE_ATOMIC_MASK)
1011                 send_atomic_ack(qp, pkt, AETH_ACK_UNLIMITED);
1012         else if (bth_ack(pkt))
1013                 send_ack(qp, pkt, AETH_ACK_UNLIMITED, pkt->psn);
1014
1015         return RESPST_CLEANUP;
1016 }
1017
1018 static enum resp_states cleanup(struct rxe_qp *qp,
1019                                 struct rxe_pkt_info *pkt)
1020 {
1021         struct sk_buff *skb;
1022
1023         if (pkt) {
1024                 skb = skb_dequeue(&qp->req_pkts);
1025                 rxe_drop_ref(qp);
1026                 kfree_skb(skb);
1027         }
1028
1029         if (qp->resp.mr) {
1030                 rxe_drop_ref(qp->resp.mr);
1031                 qp->resp.mr = NULL;
1032         }
1033
1034         return RESPST_DONE;
1035 }
1036
1037 static struct resp_res *find_resource(struct rxe_qp *qp, u32 psn)
1038 {
1039         int i;
1040
1041         for (i = 0; i < qp->attr.max_rd_atomic; i++) {
1042                 struct resp_res *res = &qp->resp.resources[i];
1043
1044                 if (res->type == 0)
1045                         continue;
1046
1047                 if (psn_compare(psn, res->first_psn) >= 0 &&
1048                     psn_compare(psn, res->last_psn) <= 0) {
1049                         return res;
1050                 }
1051         }
1052
1053         return NULL;
1054 }
1055
1056 static enum resp_states duplicate_request(struct rxe_qp *qp,
1057                                           struct rxe_pkt_info *pkt)
1058 {
1059         enum resp_states rc;
1060
1061         if (pkt->mask & RXE_SEND_MASK ||
1062             pkt->mask & RXE_WRITE_MASK) {
1063                 /* SEND. Ack again and cleanup. C9-105. */
1064                 if (bth_ack(pkt))
1065                         send_ack(qp, pkt, AETH_ACK_UNLIMITED, qp->resp.psn - 1);
1066                 rc = RESPST_CLEANUP;
1067                 goto out;
1068         } else if (pkt->mask & RXE_READ_MASK) {
1069                 struct resp_res *res;
1070
1071                 res = find_resource(qp, pkt->psn);
1072                 if (!res) {
1073                         /* Resource not found. Class D error.  Drop the
1074                          * request.
1075                          */
1076                         rc = RESPST_CLEANUP;
1077                         goto out;
1078                 } else {
1079                         /* Ensure this new request is the same as the previous
1080                          * one or a subset of it.
1081                          */
1082                         u64 iova = reth_va(pkt);
1083                         u32 resid = reth_len(pkt);
1084
1085                         if (iova < res->read.va_org ||
1086                             resid > res->read.length ||
1087                             (iova + resid) > (res->read.va_org +
1088                                               res->read.length)) {
1089                                 rc = RESPST_CLEANUP;
1090                                 goto out;
1091                         }
1092
1093                         if (reth_rkey(pkt) != res->read.rkey) {
1094                                 rc = RESPST_CLEANUP;
1095                                 goto out;
1096                         }
1097
1098                         res->cur_psn = pkt->psn;
1099                         res->state = (pkt->psn == res->first_psn) ?
1100                                         rdatm_res_state_new :
1101                                         rdatm_res_state_replay;
1102
1103                         /* Reset the resource, except length. */
1104                         res->read.va_org = iova;
1105                         res->read.va = iova;
1106                         res->read.resid = resid;
1107
1108                         /* Replay the RDMA read reply. */
1109                         qp->resp.res = res;
1110                         rc = RESPST_READ_REPLY;
1111                         goto out;
1112                 }
1113         } else {
1114                 struct resp_res *res;
1115
1116                 /* Find the operation in our list of responder resources. */
1117                 res = find_resource(qp, pkt->psn);
1118                 if (res) {
1119                         struct sk_buff *skb_copy;
1120
1121                         skb_copy = skb_clone(res->atomic.skb, GFP_ATOMIC);
1122                         if (skb_copy) {
1123                                 rxe_add_ref(qp); /* for the new SKB */
1124                         } else {
1125                                 pr_warn("Couldn't clone atomic resp\n");
1126                                 rc = RESPST_CLEANUP;
1127                                 goto out;
1128                         }
1129
1130                         /* Resend the result. */
1131                         rc = rxe_xmit_packet(to_rdev(qp->ibqp.device), qp,
1132                                              pkt, skb_copy);
1133                         if (rc) {
1134                                 pr_err("Failed resending result. This flow is not handled - skb ignored\n");
1135                                 kfree_skb(skb_copy);
1136                                 rc = RESPST_CLEANUP;
1137                                 goto out;
1138                         }
1139                 }
1140
1141                 /* Resource not found. Class D error. Drop the request. */
1142                 rc = RESPST_CLEANUP;
1143                 goto out;
1144         }
1145 out:
1146         return rc;
1147 }
1148
1149 /* Process a class A or C. Both are treated the same in this implementation. */
1150 static void do_class_ac_error(struct rxe_qp *qp, u8 syndrome,
1151                               enum ib_wc_status status)
1152 {
1153         qp->resp.aeth_syndrome  = syndrome;
1154         qp->resp.status         = status;
1155
1156         /* indicate that we should go through the ERROR state */
1157         qp->resp.goto_error     = 1;
1158 }
1159
1160 static enum resp_states do_class_d1e_error(struct rxe_qp *qp)
1161 {
1162         /* UC */
1163         if (qp->srq) {
1164                 /* Class E */
1165                 qp->resp.drop_msg = 1;
1166                 if (qp->resp.wqe) {
1167                         qp->resp.status = IB_WC_REM_INV_REQ_ERR;
1168                         return RESPST_COMPLETE;
1169                 } else {
1170                         return RESPST_CLEANUP;
1171                 }
1172         } else {
1173                 /* Class D1. This packet may be the start of a
1174                  * new message and could be valid. The previous
1175                  * message is invalid and ignored. reset the
1176                  * recv wr to its original state
1177                  */
1178                 if (qp->resp.wqe) {
1179                         qp->resp.wqe->dma.resid = qp->resp.wqe->dma.length;
1180                         qp->resp.wqe->dma.cur_sge = 0;
1181                         qp->resp.wqe->dma.sge_offset = 0;
1182                         qp->resp.opcode = -1;
1183                 }
1184
1185                 if (qp->resp.mr) {
1186                         rxe_drop_ref(qp->resp.mr);
1187                         qp->resp.mr = NULL;
1188                 }
1189
1190                 return RESPST_CLEANUP;
1191         }
1192 }
1193
1194 int rxe_responder(void *arg)
1195 {
1196         struct rxe_qp *qp = (struct rxe_qp *)arg;
1197         enum resp_states state;
1198         struct rxe_pkt_info *pkt = NULL;
1199         int ret = 0;
1200
1201         qp->resp.aeth_syndrome = AETH_ACK_UNLIMITED;
1202
1203         if (!qp->valid) {
1204                 ret = -EINVAL;
1205                 goto done;
1206         }
1207
1208         switch (qp->resp.state) {
1209         case QP_STATE_RESET:
1210                 state = RESPST_RESET;
1211                 break;
1212
1213         default:
1214                 state = RESPST_GET_REQ;
1215                 break;
1216         }
1217
1218         while (1) {
1219                 pr_debug("qp#%d state = %s\n", qp_num(qp),
1220                          resp_state_name[state]);
1221                 switch (state) {
1222                 case RESPST_GET_REQ:
1223                         state = get_req(qp, &pkt);
1224                         break;
1225                 case RESPST_CHK_PSN:
1226                         state = check_psn(qp, pkt);
1227                         break;
1228                 case RESPST_CHK_OP_SEQ:
1229                         state = check_op_seq(qp, pkt);
1230                         break;
1231                 case RESPST_CHK_OP_VALID:
1232                         state = check_op_valid(qp, pkt);
1233                         break;
1234                 case RESPST_CHK_RESOURCE:
1235                         state = check_resource(qp, pkt);
1236                         break;
1237                 case RESPST_CHK_LENGTH:
1238                         state = check_length(qp, pkt);
1239                         break;
1240                 case RESPST_CHK_RKEY:
1241                         state = check_rkey(qp, pkt);
1242                         break;
1243                 case RESPST_EXECUTE:
1244                         state = execute(qp, pkt);
1245                         break;
1246                 case RESPST_COMPLETE:
1247                         state = do_complete(qp, pkt);
1248                         break;
1249                 case RESPST_READ_REPLY:
1250                         state = read_reply(qp, pkt);
1251                         break;
1252                 case RESPST_ACKNOWLEDGE:
1253                         state = acknowledge(qp, pkt);
1254                         break;
1255                 case RESPST_CLEANUP:
1256                         state = cleanup(qp, pkt);
1257                         break;
1258                 case RESPST_DUPLICATE_REQUEST:
1259                         state = duplicate_request(qp, pkt);
1260                         break;
1261                 case RESPST_ERR_PSN_OUT_OF_SEQ:
1262                         /* RC only - Class B. Drop packet. */
1263                         send_ack(qp, pkt, AETH_NAK_PSN_SEQ_ERROR, qp->resp.psn);
1264                         state = RESPST_CLEANUP;
1265                         break;
1266
1267                 case RESPST_ERR_TOO_MANY_RDMA_ATM_REQ:
1268                 case RESPST_ERR_MISSING_OPCODE_FIRST:
1269                 case RESPST_ERR_MISSING_OPCODE_LAST_C:
1270                 case RESPST_ERR_UNSUPPORTED_OPCODE:
1271                 case RESPST_ERR_MISALIGNED_ATOMIC:
1272                         /* RC Only - Class C. */
1273                         do_class_ac_error(qp, AETH_NAK_INVALID_REQ,
1274                                           IB_WC_REM_INV_REQ_ERR);
1275                         state = RESPST_COMPLETE;
1276                         break;
1277
1278                 case RESPST_ERR_MISSING_OPCODE_LAST_D1E:
1279                         state = do_class_d1e_error(qp);
1280                         break;
1281                 case RESPST_ERR_RNR:
1282                         if (qp_type(qp) == IB_QPT_RC) {
1283                                 /* RC - class B */
1284                                 send_ack(qp, pkt, AETH_RNR_NAK |
1285                                          (~AETH_TYPE_MASK &
1286                                          qp->attr.min_rnr_timer),
1287                                          pkt->psn);
1288                         } else {
1289                                 /* UD/UC - class D */
1290                                 qp->resp.drop_msg = 1;
1291                         }
1292                         state = RESPST_CLEANUP;
1293                         break;
1294
1295                 case RESPST_ERR_RKEY_VIOLATION:
1296                         if (qp_type(qp) == IB_QPT_RC) {
1297                                 /* Class C */
1298                                 do_class_ac_error(qp, AETH_NAK_REM_ACC_ERR,
1299                                                   IB_WC_REM_ACCESS_ERR);
1300                                 state = RESPST_COMPLETE;
1301                         } else {
1302                                 qp->resp.drop_msg = 1;
1303                                 if (qp->srq) {
1304                                         /* UC/SRQ Class D */
1305                                         qp->resp.status = IB_WC_REM_ACCESS_ERR;
1306                                         state = RESPST_COMPLETE;
1307                                 } else {
1308                                         /* UC/non-SRQ Class E. */
1309                                         state = RESPST_CLEANUP;
1310                                 }
1311                         }
1312                         break;
1313
1314                 case RESPST_ERR_LENGTH:
1315                         if (qp_type(qp) == IB_QPT_RC) {
1316                                 /* Class C */
1317                                 do_class_ac_error(qp, AETH_NAK_INVALID_REQ,
1318                                                   IB_WC_REM_INV_REQ_ERR);
1319                                 state = RESPST_COMPLETE;
1320                         } else if (qp->srq) {
1321                                 /* UC/UD - class E */
1322                                 qp->resp.status = IB_WC_REM_INV_REQ_ERR;
1323                                 state = RESPST_COMPLETE;
1324                         } else {
1325                                 /* UC/UD - class D */
1326                                 qp->resp.drop_msg = 1;
1327                                 state = RESPST_CLEANUP;
1328                         }
1329                         break;
1330
1331                 case RESPST_ERR_MALFORMED_WQE:
1332                         /* All, Class A. */
1333                         do_class_ac_error(qp, AETH_NAK_REM_OP_ERR,
1334                                           IB_WC_LOC_QP_OP_ERR);
1335                         state = RESPST_COMPLETE;
1336                         break;
1337
1338                 case RESPST_ERR_CQ_OVERFLOW:
1339                         /* All - Class G */
1340                         state = RESPST_ERROR;
1341                         break;
1342
1343                 case RESPST_DONE:
1344                         if (qp->resp.goto_error) {
1345                                 state = RESPST_ERROR;
1346                                 break;
1347                         }
1348
1349                         goto done;
1350
1351                 case RESPST_EXIT:
1352                         if (qp->resp.goto_error) {
1353                                 state = RESPST_ERROR;
1354                                 break;
1355                         }
1356
1357                         goto exit;
1358
1359                 case RESPST_RESET: {
1360                         struct sk_buff *skb;
1361
1362                         while ((skb = skb_dequeue(&qp->req_pkts))) {
1363                                 rxe_drop_ref(qp);
1364                                 kfree_skb(skb);
1365                         }
1366
1367                         while (!qp->srq && qp->rq.queue &&
1368                                queue_head(qp->rq.queue))
1369                                 advance_consumer(qp->rq.queue);
1370
1371                         qp->resp.wqe = NULL;
1372                         goto exit;
1373                 }
1374
1375                 case RESPST_ERROR:
1376                         qp->resp.goto_error = 0;
1377                         pr_warn("qp#%d moved to error state\n", qp_num(qp));
1378                         rxe_qp_error(qp);
1379                         goto exit;
1380
1381                 default:
1382                         WARN_ON(1);
1383                 }
1384         }
1385
1386 exit:
1387         ret = -EAGAIN;
1388 done:
1389         return ret;
1390 }