drivers/infiniband/hw/ipath/ipath_rc.c

   1 /*
   2  * Copyright (c) 2006 QLogic, Inc. All rights reserved.
   3  * Copyright (c) 2005, 2006 PathScale, 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 "ipath_verbs.h"
  35 #include "ipath_kernel.h"
  36
  37 /* cut down ridiculously long IB macro names */
  38 #define OP(x) IB_OPCODE_RC_##x
  39
  40 static u32 restart_sge(struct ipath_sge_state *ss, struct ipath_swqe *wqe,
  41                        u32 psn, u32 pmtu)
  42 {
  43         u32 len;
  44
  45         len = ((psn - wqe->psn) & IPATH_PSN_MASK) * pmtu;
  46         ss->sge = wqe->sg_list[0];
  47         ss->sg_list = wqe->sg_list + 1;
  48         ss->num_sge = wqe->wr.num_sge;
  49         ipath_skip_sge(ss, len);
  50         return wqe->length - len;
  51 }
  52
  53 /**
  54  * ipath_init_restart- initialize the qp->s_sge after a restart
  55  * @qp: the QP who's SGE we're restarting
  56  * @wqe: the work queue to initialize the QP's SGE from
  57  *
  58  * The QP s_lock should be held and interrupts disabled.
  59  */
  60 static void ipath_init_restart(struct ipath_qp *qp, struct ipath_swqe *wqe)
  61 {
  62         struct ipath_ibdev *dev;
  63
  64         qp->s_len = restart_sge(&qp->s_sge, wqe, qp->s_psn,
  65                                 ib_mtu_enum_to_int(qp->path_mtu));
  66         dev = to_idev(qp->ibqp.device);
  67         spin_lock(&dev->pending_lock);
  68         if (list_empty(&qp->timerwait))
  69                 list_add_tail(&qp->timerwait,
  70                               &dev->pending[dev->pending_index]);
  71         spin_unlock(&dev->pending_lock);
  72 }
  73
  74 /**
  75  * ipath_make_rc_ack - construct a response packet (ACK, NAK, or RDMA read)
  76  * @qp: a pointer to the QP
  77  * @ohdr: a pointer to the IB header being constructed
  78  * @pmtu: the path MTU
  79  *
  80  * Return 1 if constructed; otherwise, return 0.
  81  * Note that we are in the responder's side of the QP context.
  82  * Note the QP s_lock must be held.
  83  */
  84 static int ipath_make_rc_ack(struct ipath_qp *qp,
  85                              struct ipath_other_headers *ohdr,
  86                              u32 pmtu, u32 *bth0p, u32 *bth2p)
  87 {
  88         struct ipath_ack_entry *e;
  89         u32 hwords;
  90         u32 len;
  91         u32 bth0;
  92         u32 bth2;
  93
  94         /* header size in 32-bit words LRH+BTH = (8+12)/4. */
  95         hwords = 5;
  96
  97         switch (qp->s_ack_state) {
  98         case OP(RDMA_READ_RESPONSE_LAST):
  99         case OP(RDMA_READ_RESPONSE_ONLY):
 100         case OP(ATOMIC_ACKNOWLEDGE):
 101                 qp->s_ack_state = OP(ACKNOWLEDGE);
 102                 /* FALLTHROUGH */
 103         case OP(ACKNOWLEDGE):
 104                 /* Check for no next entry in the queue. */
 105                 if (qp->r_head_ack_queue == qp->s_tail_ack_queue) {
 106                         if (qp->s_flags & IPATH_S_ACK_PENDING)
 107                                 goto normal;
 108                         goto bail;
 109                 }
 110
 111                 e = &qp->s_ack_queue[qp->s_tail_ack_queue];
 112                 if (e->opcode == OP(RDMA_READ_REQUEST)) {
 113                         /* Copy SGE state in case we need to resend */
 114                         qp->s_ack_rdma_sge = e->rdma_sge;
 115                         qp->s_cur_sge = &qp->s_ack_rdma_sge;
 116                         len = e->rdma_sge.sge.sge_length;
 117                         if (len > pmtu) {
 118                                 len = pmtu;
 119                                 qp->s_ack_state = OP(RDMA_READ_RESPONSE_FIRST);
 120                         } else {
 121                                 qp->s_ack_state = OP(RDMA_READ_RESPONSE_ONLY);
 122                                 if (++qp->s_tail_ack_queue >
 123                                     IPATH_MAX_RDMA_ATOMIC)
 124                                         qp->s_tail_ack_queue = 0;
 125                         }
 126                         ohdr->u.aeth = ipath_compute_aeth(qp);
 127                         hwords++;
 128                         qp->s_ack_rdma_psn = e->psn;
 129                         bth2 = qp->s_ack_rdma_psn++ & IPATH_PSN_MASK;
 130                 } else {
 131                         /* COMPARE_SWAP or FETCH_ADD */
 132                         qp->s_cur_sge = NULL;
 133                         len = 0;
 134                         qp->s_ack_state = OP(ATOMIC_ACKNOWLEDGE);
 135                         ohdr->u.at.aeth = ipath_compute_aeth(qp);
 136                         ohdr->u.at.atomic_ack_eth[0] =
 137                                 cpu_to_be32(e->atomic_data >> 32);
 138                         ohdr->u.at.atomic_ack_eth[1] =
 139                                 cpu_to_be32(e->atomic_data);
 140                         hwords += sizeof(ohdr->u.at) / sizeof(u32);
 141                         bth2 = e->psn;
 142                         if (++qp->s_tail_ack_queue > IPATH_MAX_RDMA_ATOMIC)
 143                                 qp->s_tail_ack_queue = 0;
 144                 }
 145                 bth0 = qp->s_ack_state << 24;
 146                 break;
 147
 148         case OP(RDMA_READ_RESPONSE_FIRST):
 149                 qp->s_ack_state = OP(RDMA_READ_RESPONSE_MIDDLE);
 150                 /* FALLTHROUGH */
 151         case OP(RDMA_READ_RESPONSE_MIDDLE):
 152                 len = qp->s_ack_rdma_sge.sge.sge_length;
 153                 if (len > pmtu)
 154                         len = pmtu;
 155                 else {
 156                         ohdr->u.aeth = ipath_compute_aeth(qp);
 157                         hwords++;
 158                         qp->s_ack_state = OP(RDMA_READ_RESPONSE_LAST);
 159                         if (++qp->s_tail_ack_queue > IPATH_MAX_RDMA_ATOMIC)
 160                                 qp->s_tail_ack_queue = 0;
 161                 }
 162                 bth0 = qp->s_ack_state << 24;
 163                 bth2 = qp->s_ack_rdma_psn++ & IPATH_PSN_MASK;
 164                 break;
 165
 166         default:
 167         normal:
 168                 /*
 169                  * Send a regular ACK.
 170                  * Set the s_ack_state so we wait until after sending
 171                  * the ACK before setting s_ack_state to ACKNOWLEDGE
 172                  * (see above).
 173                  */
 174                 qp->s_ack_state = OP(ATOMIC_ACKNOWLEDGE);
 175                 qp->s_flags &= ~IPATH_S_ACK_PENDING;
 176                 qp->s_cur_sge = NULL;
 177                 if (qp->s_nak_state)
 178                         ohdr->u.aeth =
 179                                 cpu_to_be32((qp->r_msn & IPATH_MSN_MASK) |
 180                                             (qp->s_nak_state <<
 181                                              IPATH_AETH_CREDIT_SHIFT));
 182                 else
 183                         ohdr->u.aeth = ipath_compute_aeth(qp);
 184                 hwords++;
 185                 len = 0;
 186                 bth0 = OP(ACKNOWLEDGE) << 24;
 187                 bth2 = qp->s_ack_psn & IPATH_PSN_MASK;
 188         }
 189         qp->s_hdrwords = hwords;
 190         qp->s_cur_size = len;
 191         *bth0p = bth0;
 192         *bth2p = bth2;
 193         return 1;
 194
 195 bail:
 196         return 0;
 197 }
 198
 199 /**
 200  * ipath_make_rc_req - construct a request packet (SEND, RDMA r/w, ATOMIC)
 201  * @qp: a pointer to the QP
 202  * @ohdr: a pointer to the IB header being constructed
 203  * @pmtu: the path MTU
 204  * @bth0p: pointer to the BTH opcode word
 205  * @bth2p: pointer to the BTH PSN word
 206  *
 207  * Return 1 if constructed; otherwise, return 0.
 208  * Note the QP s_lock must be held and interrupts disabled.
 209  */
 210 int ipath_make_rc_req(struct ipath_qp *qp,
 211                       struct ipath_other_headers *ohdr,
 212                       u32 pmtu, u32 *bth0p, u32 *bth2p)
 213 {
 214         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
 215         struct ipath_sge_state *ss;
 216         struct ipath_swqe *wqe;
 217         u32 hwords;
 218         u32 len;
 219         u32 bth0;
 220         u32 bth2;
 221         char newreq;
 222
 223         /* Sending responses has higher priority over sending requests. */
 224         if ((qp->r_head_ack_queue != qp->s_tail_ack_queue ||
 225              (qp->s_flags & IPATH_S_ACK_PENDING) ||
 226              qp->s_ack_state != IB_OPCODE_RC_ACKNOWLEDGE) &&
 227             ipath_make_rc_ack(qp, ohdr, pmtu, bth0p, bth2p))
 228                 goto done;
 229
 230         if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK) ||
 231             qp->s_rnr_timeout)
 232                 goto bail;
 233
 234         /* Limit the number of packets sent without an ACK. */
 235         if (ipath_cmp24(qp->s_psn, qp->s_last_psn + IPATH_PSN_CREDIT) > 0) {
 236                 qp->s_wait_credit = 1;
 237                 dev->n_rc_stalls++;
 238                 spin_lock(&dev->pending_lock);
 239                 if (list_empty(&qp->timerwait))
 240                         list_add_tail(&qp->timerwait,
 241                                       &dev->pending[dev->pending_index]);
 242                 spin_unlock(&dev->pending_lock);
 243                 goto bail;
 244         }
 245
 246         /* header size in 32-bit words LRH+BTH = (8+12)/4. */
 247         hwords = 5;
 248         bth0 = 0;
 249
 250         /* Send a request. */
 251         wqe = get_swqe_ptr(qp, qp->s_cur);
 252         switch (qp->s_state) {
 253         default:
 254                 /*
 255                  * Resend an old request or start a new one.
 256                  *
 257                  * We keep track of the current SWQE so that
 258                  * we don't reset the "furthest progress" state
 259                  * if we need to back up.
 260                  */
 261                 newreq = 0;
 262                 if (qp->s_cur == qp->s_tail) {
 263                         /* Check if send work queue is empty. */
 264                         if (qp->s_tail == qp->s_head)
 265                                 goto bail;
 266                         /*
 267                          * If a fence is requested, wait for previous
 268                          * RDMA read and atomic operations to finish.
 269                          */
 270                         if ((wqe->wr.send_flags & IB_SEND_FENCE) &&
 271                             qp->s_num_rd_atomic) {
 272                                 qp->s_flags |= IPATH_S_FENCE_PENDING;
 273                                 goto bail;
 274                         }
 275                         wqe->psn = qp->s_next_psn;
 276                         newreq = 1;
 277                 }
 278                 /*
 279                  * Note that we have to be careful not to modify the
 280                  * original work request since we may need to resend
 281                  * it.
 282                  */
 283                 len = wqe->length;
 284                 ss = &qp->s_sge;
 285                 bth2 = 0;
 286                 switch (wqe->wr.opcode) {
 287                 case IB_WR_SEND:
 288                 case IB_WR_SEND_WITH_IMM:
 289                         /* If no credit, return. */
 290                         if (qp->s_lsn != (u32) -1 &&
 291                             ipath_cmp24(wqe->ssn, qp->s_lsn + 1) > 0)
 292                                 goto bail;
 293                         wqe->lpsn = wqe->psn;
 294                         if (len > pmtu) {
 295                                 wqe->lpsn += (len - 1) / pmtu;
 296                                 qp->s_state = OP(SEND_FIRST);
 297                                 len = pmtu;
 298                                 break;
 299                         }
 300                         if (wqe->wr.opcode == IB_WR_SEND)
 301                                 qp->s_state = OP(SEND_ONLY);
 302                         else {
 303                                 qp->s_state = OP(SEND_ONLY_WITH_IMMEDIATE);
 304                                 /* Immediate data comes after the BTH */
 305                                 ohdr->u.imm_data = wqe->wr.imm_data;
 306                                 hwords += 1;
 307                         }
 308                         if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 309                                 bth0 |= 1 << 23;
 310                         bth2 = 1 << 31; /* Request ACK. */
 311                         if (++qp->s_cur == qp->s_size)
 312                                 qp->s_cur = 0;
 313                         break;
 314
 315                 case IB_WR_RDMA_WRITE:
 316                         if (newreq && qp->s_lsn != (u32) -1)
 317                                 qp->s_lsn++;
 318                         /* FALLTHROUGH */
 319                 case IB_WR_RDMA_WRITE_WITH_IMM:
 320                         /* If no credit, return. */
 321                         if (qp->s_lsn != (u32) -1 &&
 322                             ipath_cmp24(wqe->ssn, qp->s_lsn + 1) > 0)
 323                                 goto bail;
 324                         ohdr->u.rc.reth.vaddr =
 325                                 cpu_to_be64(wqe->wr.wr.rdma.remote_addr);
 326                         ohdr->u.rc.reth.rkey =
 327                                 cpu_to_be32(wqe->wr.wr.rdma.rkey);
 328                         ohdr->u.rc.reth.length = cpu_to_be32(len);
 329                         hwords += sizeof(struct ib_reth) / sizeof(u32);
 330                         wqe->lpsn = wqe->psn;
 331                         if (len > pmtu) {
 332                                 wqe->lpsn += (len - 1) / pmtu;
 333                                 qp->s_state = OP(RDMA_WRITE_FIRST);
 334                                 len = pmtu;
 335                                 break;
 336                         }
 337                         if (wqe->wr.opcode == IB_WR_RDMA_WRITE)
 338                                 qp->s_state = OP(RDMA_WRITE_ONLY);
 339                         else {
 340                                 qp->s_state =
 341                                         OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE);
 342                                 /* Immediate data comes after RETH */
 343                                 ohdr->u.rc.imm_data = wqe->wr.imm_data;
 344                                 hwords += 1;
 345                                 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 346                                         bth0 |= 1 << 23;
 347                         }
 348                         bth2 = 1 << 31; /* Request ACK. */
 349                         if (++qp->s_cur == qp->s_size)
 350                                 qp->s_cur = 0;
 351                         break;
 352
 353                 case IB_WR_RDMA_READ:
 354                         /*
 355                          * Don't allow more operations to be started
 356                          * than the QP limits allow.
 357                          */
 358                         if (newreq) {
 359                                 if (qp->s_num_rd_atomic >=
 360                                     qp->s_max_rd_atomic) {
 361                                         qp->s_flags |= IPATH_S_RDMAR_PENDING;
 362                                         goto bail;
 363                                 }
 364                                 qp->s_num_rd_atomic++;
 365                                 if (qp->s_lsn != (u32) -1)
 366                                         qp->s_lsn++;
 367                                 /*
 368                                  * Adjust s_next_psn to count the
 369                                  * expected number of responses.
 370                                  */
 371                                 if (len > pmtu)
 372                                         qp->s_next_psn += (len - 1) / pmtu;
 373                                 wqe->lpsn = qp->s_next_psn++;
 374                         }
 375                         ohdr->u.rc.reth.vaddr =
 376                                 cpu_to_be64(wqe->wr.wr.rdma.remote_addr);
 377                         ohdr->u.rc.reth.rkey =
 378                                 cpu_to_be32(wqe->wr.wr.rdma.rkey);
 379                         ohdr->u.rc.reth.length = cpu_to_be32(len);
 380                         qp->s_state = OP(RDMA_READ_REQUEST);
 381                         hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
 382                         ss = NULL;
 383                         len = 0;
 384                         if (++qp->s_cur == qp->s_size)
 385                                 qp->s_cur = 0;
 386                         break;
 387
 388                 case IB_WR_ATOMIC_CMP_AND_SWP:
 389                 case IB_WR_ATOMIC_FETCH_AND_ADD:
 390                         /*
 391                          * Don't allow more operations to be started
 392                          * than the QP limits allow.
 393                          */
 394                         if (newreq) {
 395                                 if (qp->s_num_rd_atomic >=
 396                                     qp->s_max_rd_atomic) {
 397                                         qp->s_flags |= IPATH_S_RDMAR_PENDING;
 398                                         goto bail;
 399                                 }
 400                                 qp->s_num_rd_atomic++;
 401                                 if (qp->s_lsn != (u32) -1)
 402                                         qp->s_lsn++;
 403                                 wqe->lpsn = wqe->psn;
 404                         }
 405                         if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
 406                                 qp->s_state = OP(COMPARE_SWAP);
 407                                 ohdr->u.atomic_eth.swap_data = cpu_to_be64(
 408                                         wqe->wr.wr.atomic.swap);
 409                                 ohdr->u.atomic_eth.compare_data = cpu_to_be64(
 410                                         wqe->wr.wr.atomic.compare_add);
 411                         } else {
 412                                 qp->s_state = OP(FETCH_ADD);
 413                                 ohdr->u.atomic_eth.swap_data = cpu_to_be64(
 414                                         wqe->wr.wr.atomic.compare_add);
 415                                 ohdr->u.atomic_eth.compare_data = 0;
 416                         }
 417                         ohdr->u.atomic_eth.vaddr[0] = cpu_to_be32(
 418                                 wqe->wr.wr.atomic.remote_addr >> 32);
 419                         ohdr->u.atomic_eth.vaddr[1] = cpu_to_be32(
 420                                 wqe->wr.wr.atomic.remote_addr);
 421                         ohdr->u.atomic_eth.rkey = cpu_to_be32(
 422                                 wqe->wr.wr.atomic.rkey);
 423                         hwords += sizeof(struct ib_atomic_eth) / sizeof(u32);
 424                         ss = NULL;
 425                         len = 0;
 426                         if (++qp->s_cur == qp->s_size)
 427                                 qp->s_cur = 0;
 428                         break;
 429
 430                 default:
 431                         goto bail;
 432                 }
 433                 qp->s_sge.sge = wqe->sg_list[0];
 434                 qp->s_sge.sg_list = wqe->sg_list + 1;
 435                 qp->s_sge.num_sge = wqe->wr.num_sge;
 436                 qp->s_len = wqe->length;
 437                 if (newreq) {
 438                         qp->s_tail++;
 439                         if (qp->s_tail >= qp->s_size)
 440                                 qp->s_tail = 0;
 441                 }
 442                 bth2 |= qp->s_psn & IPATH_PSN_MASK;
 443                 if (wqe->wr.opcode == IB_WR_RDMA_READ)
 444                         qp->s_psn = wqe->lpsn + 1;
 445                 else {
 446                         qp->s_psn++;
 447                         if (ipath_cmp24(qp->s_psn, qp->s_next_psn) > 0)
 448                                 qp->s_next_psn = qp->s_psn;
 449                 }
 450                 /*
 451                  * Put the QP on the pending list so lost ACKs will cause
 452                  * a retry.  More than one request can be pending so the
 453                  * QP may already be on the dev->pending list.
 454                  */
 455                 spin_lock(&dev->pending_lock);
 456                 if (list_empty(&qp->timerwait))
 457                         list_add_tail(&qp->timerwait,
 458                                       &dev->pending[dev->pending_index]);
 459                 spin_unlock(&dev->pending_lock);
 460                 break;
 461
 462         case OP(RDMA_READ_RESPONSE_FIRST):
 463                 /*
 464                  * This case can only happen if a send is restarted.
 465                  * See ipath_restart_rc().
 466                  */
 467                 ipath_init_restart(qp, wqe);
 468                 /* FALLTHROUGH */
 469         case OP(SEND_FIRST):
 470                 qp->s_state = OP(SEND_MIDDLE);
 471                 /* FALLTHROUGH */
 472         case OP(SEND_MIDDLE):
 473                 bth2 = qp->s_psn++ & IPATH_PSN_MASK;
 474                 if (ipath_cmp24(qp->s_psn, qp->s_next_psn) > 0)
 475                         qp->s_next_psn = qp->s_psn;
 476                 ss = &qp->s_sge;
 477                 len = qp->s_len;
 478                 if (len > pmtu) {
 479                         len = pmtu;
 480                         break;
 481                 }
 482                 if (wqe->wr.opcode == IB_WR_SEND)
 483                         qp->s_state = OP(SEND_LAST);
 484                 else {
 485                         qp->s_state = OP(SEND_LAST_WITH_IMMEDIATE);
 486                         /* Immediate data comes after the BTH */
 487                         ohdr->u.imm_data = wqe->wr.imm_data;
 488                         hwords += 1;
 489                 }
 490                 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 491                         bth0 |= 1 << 23;
 492                 bth2 |= 1 << 31;        /* Request ACK. */
 493                 qp->s_cur++;
 494                 if (qp->s_cur >= qp->s_size)
 495                         qp->s_cur = 0;
 496                 break;
 497
 498         case OP(RDMA_READ_RESPONSE_LAST):
 499                 /*
 500                  * This case can only happen if a RDMA write is restarted.
 501                  * See ipath_restart_rc().
 502                  */
 503                 ipath_init_restart(qp, wqe);
 504                 /* FALLTHROUGH */
 505         case OP(RDMA_WRITE_FIRST):
 506                 qp->s_state = OP(RDMA_WRITE_MIDDLE);
 507                 /* FALLTHROUGH */
 508         case OP(RDMA_WRITE_MIDDLE):
 509                 bth2 = qp->s_psn++ & IPATH_PSN_MASK;
 510                 if (ipath_cmp24(qp->s_psn, qp->s_next_psn) > 0)
 511                         qp->s_next_psn = qp->s_psn;
 512                 ss = &qp->s_sge;
 513                 len = qp->s_len;
 514                 if (len > pmtu) {
 515                         len = pmtu;
 516                         break;
 517                 }
 518                 if (wqe->wr.opcode == IB_WR_RDMA_WRITE)
 519                         qp->s_state = OP(RDMA_WRITE_LAST);
 520                 else {
 521                         qp->s_state = OP(RDMA_WRITE_LAST_WITH_IMMEDIATE);
 522                         /* Immediate data comes after the BTH */
 523                         ohdr->u.imm_data = wqe->wr.imm_data;
 524                         hwords += 1;
 525                         if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 526                                 bth0 |= 1 << 23;
 527                 }
 528                 bth2 |= 1 << 31;        /* Request ACK. */
 529                 qp->s_cur++;
 530                 if (qp->s_cur >= qp->s_size)
 531                         qp->s_cur = 0;
 532                 break;
 533
 534         case OP(RDMA_READ_RESPONSE_MIDDLE):
 535                 /*
 536                  * This case can only happen if a RDMA read is restarted.
 537                  * See ipath_restart_rc().
 538                  */
 539                 ipath_init_restart(qp, wqe);
 540                 len = ((qp->s_psn - wqe->psn) & IPATH_PSN_MASK) * pmtu;
 541                 ohdr->u.rc.reth.vaddr =
 542                         cpu_to_be64(wqe->wr.wr.rdma.remote_addr + len);
 543                 ohdr->u.rc.reth.rkey =
 544                         cpu_to_be32(wqe->wr.wr.rdma.rkey);
 545                 ohdr->u.rc.reth.length = cpu_to_be32(qp->s_len);
 546                 qp->s_state = OP(RDMA_READ_REQUEST);
 547                 hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
 548                 bth2 = qp->s_psn++ & IPATH_PSN_MASK;
 549                 if (ipath_cmp24(qp->s_psn, qp->s_next_psn) > 0)
 550                         qp->s_next_psn = qp->s_psn;
 551                 ss = NULL;
 552                 len = 0;
 553                 qp->s_cur++;
 554                 if (qp->s_cur == qp->s_size)
 555                         qp->s_cur = 0;
 556                 break;
 557         }
 558         if (ipath_cmp24(qp->s_psn, qp->s_last_psn + IPATH_PSN_CREDIT - 1) >= 0)
 559                 bth2 |= 1 << 31;        /* Request ACK. */
 560         qp->s_len -= len;
 561         qp->s_hdrwords = hwords;
 562         qp->s_cur_sge = ss;
 563         qp->s_cur_size = len;
 564         *bth0p = bth0 | (qp->s_state << 24);
 565         *bth2p = bth2;
 566 done:
 567         return 1;
 568
 569 bail:
 570         return 0;
 571 }
 572
 573 /**
 574  * send_rc_ack - Construct an ACK packet and send it
 575  * @qp: a pointer to the QP
 576  *
 577  * This is called from ipath_rc_rcv() and only uses the receive
 578  * side QP state.
 579  * Note that RDMA reads and atomics are handled in the
 580  * send side QP state and tasklet.
 581  */
 582 static void send_rc_ack(struct ipath_qp *qp)
 583 {
 584         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
 585         u16 lrh0;
 586         u32 bth0;
 587         u32 hwords;
 588         struct ipath_ib_header hdr;
 589         struct ipath_other_headers *ohdr;
 590         unsigned long flags;
 591
 592         /* Don't send ACK or NAK if a RDMA read or atomic is pending. */
 593         if (qp->r_head_ack_queue != qp->s_tail_ack_queue)
 594                 goto queue_ack;
 595
 596         /* Construct the header. */
 597         ohdr = &hdr.u.oth;
 598         lrh0 = IPATH_LRH_BTH;
 599         /* header size in 32-bit words LRH+BTH+AETH = (8+12+4)/4. */
 600         hwords = 6;
 601         if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
 602                 hwords += ipath_make_grh(dev, &hdr.u.l.grh,
 603                                          &qp->remote_ah_attr.grh,
 604                                          hwords, 0);
 605                 ohdr = &hdr.u.l.oth;
 606                 lrh0 = IPATH_LRH_GRH;
 607         }
 608         /* read pkey_index w/o lock (its atomic) */
 609         bth0 = ipath_get_pkey(dev->dd, qp->s_pkey_index) |
 610                 OP(ACKNOWLEDGE) << 24;
 611         if (qp->r_nak_state)
 612                 ohdr->u.aeth = cpu_to_be32((qp->r_msn & IPATH_MSN_MASK) |
 613                                             (qp->r_nak_state <<
 614                                              IPATH_AETH_CREDIT_SHIFT));
 615         else
 616                 ohdr->u.aeth = ipath_compute_aeth(qp);
 617         lrh0 |= qp->remote_ah_attr.sl << 4;
 618         hdr.lrh[0] = cpu_to_be16(lrh0);
 619         hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
 620         hdr.lrh[2] = cpu_to_be16(hwords + SIZE_OF_CRC);
 621         hdr.lrh[3] = cpu_to_be16(dev->dd->ipath_lid);
 622         ohdr->bth[0] = cpu_to_be32(bth0);
 623         ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
 624         ohdr->bth[2] = cpu_to_be32(qp->r_ack_psn & IPATH_PSN_MASK);
 625
 626         /*
 627          * If we can send the ACK, clear the ACK state.
 628          */
 629         if (ipath_verbs_send(dev->dd, hwords, (u32 *) &hdr, 0, NULL) == 0) {
 630                 dev->n_unicast_xmit++;
 631                 goto done;
 632         }
 633
 634         /*
 635          * We are out of PIO buffers at the moment.
 636          * Pass responsibility for sending the ACK to the
 637          * send tasklet so that when a PIO buffer becomes
 638          * available, the ACK is sent ahead of other outgoing
 639          * packets.
 640          */
 641         dev->n_rc_qacks++;
 642
 643 queue_ack:
 644         spin_lock_irqsave(&qp->s_lock, flags);
 645         qp->s_flags |= IPATH_S_ACK_PENDING;
 646         qp->s_nak_state = qp->r_nak_state;
 647         qp->s_ack_psn = qp->r_ack_psn;
 648         spin_unlock_irqrestore(&qp->s_lock, flags);
 649
 650         /* Call ipath_do_rc_send() in another thread. */
 651         tasklet_hi_schedule(&qp->s_task);
 652
 653 done:
 654         return;
 655 }
 656
 657 /**
 658  * reset_psn - reset the QP state to send starting from PSN
 659  * @qp: the QP
 660  * @psn: the packet sequence number to restart at
 661  *
 662  * This is called from ipath_rc_rcv() to process an incoming RC ACK
 663  * for the given QP.
 664  * Called at interrupt level with the QP s_lock held.
 665  */
 666 static void reset_psn(struct ipath_qp *qp, u32 psn)
 667 {
 668         u32 n = qp->s_last;
 669         struct ipath_swqe *wqe = get_swqe_ptr(qp, n);
 670         u32 opcode;
 671
 672         qp->s_cur = n;
 673
 674         /*
 675          * If we are starting the request from the beginning,
 676          * let the normal send code handle initialization.
 677          */
 678         if (ipath_cmp24(psn, wqe->psn) <= 0) {
 679                 qp->s_state = OP(SEND_LAST);
 680                 goto done;
 681         }
 682
 683         /* Find the work request opcode corresponding to the given PSN. */
 684         opcode = wqe->wr.opcode;
 685         for (;;) {
 686                 int diff;
 687
 688                 if (++n == qp->s_size)
 689                         n = 0;
 690                 if (n == qp->s_tail)
 691                         break;
 692                 wqe = get_swqe_ptr(qp, n);
 693                 diff = ipath_cmp24(psn, wqe->psn);
 694                 if (diff < 0)
 695                         break;
 696                 qp->s_cur = n;
 697                 /*
 698                  * If we are starting the request from the beginning,
 699                  * let the normal send code handle initialization.
 700                  */
 701                 if (diff == 0) {
 702                         qp->s_state = OP(SEND_LAST);
 703                         goto done;
 704                 }
 705                 opcode = wqe->wr.opcode;
 706         }
 707
 708         /*
 709          * Set the state to restart in the middle of a request.
 710          * Don't change the s_sge, s_cur_sge, or s_cur_size.
 711          * See ipath_do_rc_send().
 712          */
 713         switch (opcode) {
 714         case IB_WR_SEND:
 715         case IB_WR_SEND_WITH_IMM:
 716                 qp->s_state = OP(RDMA_READ_RESPONSE_FIRST);
 717                 break;
 718
 719         case IB_WR_RDMA_WRITE:
 720         case IB_WR_RDMA_WRITE_WITH_IMM:
 721                 qp->s_state = OP(RDMA_READ_RESPONSE_LAST);
 722                 break;
 723
 724         case IB_WR_RDMA_READ:
 725                 qp->s_state = OP(RDMA_READ_RESPONSE_MIDDLE);
 726                 break;
 727
 728         default:
 729                 /*
 730                  * This case shouldn't happen since its only
 731                  * one PSN per req.
 732                  */
 733                 qp->s_state = OP(SEND_LAST);
 734         }
 735 done:
 736         qp->s_psn = psn;
 737 }
 738
 739 /**
 740  * ipath_restart_rc - back up requester to resend the last un-ACKed request
 741  * @qp: the QP to restart
 742  * @psn: packet sequence number for the request
 743  * @wc: the work completion request
 744  *
 745  * The QP s_lock should be held and interrupts disabled.
 746  */
 747 void ipath_restart_rc(struct ipath_qp *qp, u32 psn, struct ib_wc *wc)
 748 {
 749         struct ipath_swqe *wqe = get_swqe_ptr(qp, qp->s_last);
 750         struct ipath_ibdev *dev;
 751
 752         if (qp->s_retry == 0) {
 753                 wc->wr_id = wqe->wr.wr_id;
 754                 wc->status = IB_WC_RETRY_EXC_ERR;
 755                 wc->opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
 756                 wc->vendor_err = 0;
 757                 wc->byte_len = 0;
 758                 wc->qp = &qp->ibqp;
 759                 wc->src_qp = qp->remote_qpn;
 760                 wc->pkey_index = 0;
 761                 wc->slid = qp->remote_ah_attr.dlid;
 762                 wc->sl = qp->remote_ah_attr.sl;
 763                 wc->dlid_path_bits = 0;
 764                 wc->port_num = 0;
 765                 ipath_sqerror_qp(qp, wc);
 766                 goto bail;
 767         }
 768         qp->s_retry--;
 769
 770         /*
 771          * Remove the QP from the timeout queue.
 772          * Note: it may already have been removed by ipath_ib_timer().
 773          */
 774         dev = to_idev(qp->ibqp.device);
 775         spin_lock(&dev->pending_lock);
 776         if (!list_empty(&qp->timerwait))
 777                 list_del_init(&qp->timerwait);
 778         spin_unlock(&dev->pending_lock);
 779
 780         if (wqe->wr.opcode == IB_WR_RDMA_READ)
 781                 dev->n_rc_resends++;
 782         else
 783                 dev->n_rc_resends += (qp->s_psn - psn) & IPATH_PSN_MASK;
 784
 785         reset_psn(qp, psn);
 786         tasklet_hi_schedule(&qp->s_task);
 787
 788 bail:
 789         return;
 790 }
 791
 792 static inline void update_last_psn(struct ipath_qp *qp, u32 psn)
 793 {
 794         if (qp->s_wait_credit) {
 795                 qp->s_wait_credit = 0;
 796                 tasklet_hi_schedule(&qp->s_task);
 797         }
 798         qp->s_last_psn = psn;
 799 }
 800
 801 /**
 802  * do_rc_ack - process an incoming RC ACK
 803  * @qp: the QP the ACK came in on
 804  * @psn: the packet sequence number of the ACK
 805  * @opcode: the opcode of the request that resulted in the ACK
 806  *
 807  * This is called from ipath_rc_rcv_resp() to process an incoming RC ACK
 808  * for the given QP.
 809  * Called at interrupt level with the QP s_lock held and interrupts disabled.
 810  * Returns 1 if OK, 0 if current operation should be aborted (NAK).
 811  */
 812 static int do_rc_ack(struct ipath_qp *qp, u32 aeth, u32 psn, int opcode)
 813 {
 814         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
 815         struct ib_wc wc;
 816         struct ipath_swqe *wqe;
 817         int ret = 0;
 818         u32 ack_psn;
 819
 820         /*
 821          * Remove the QP from the timeout queue (or RNR timeout queue).
 822          * If ipath_ib_timer() has already removed it,
 823          * it's OK since we hold the QP s_lock and ipath_restart_rc()
 824          * just won't find anything to restart if we ACK everything.
 825          */
 826         spin_lock(&dev->pending_lock);
 827         if (!list_empty(&qp->timerwait))
 828                 list_del_init(&qp->timerwait);
 829         spin_unlock(&dev->pending_lock);
 830
 831         /*
 832          * Note that NAKs implicitly ACK outstanding SEND and RDMA write
 833          * requests and implicitly NAK RDMA read and atomic requests issued
 834          * before the NAK'ed request.  The MSN won't include the NAK'ed
 835          * request but will include an ACK'ed request(s).
 836          */
 837         ack_psn = psn;
 838         if (aeth >> 29)
 839                 ack_psn--;
 840         wqe = get_swqe_ptr(qp, qp->s_last);
 841
 842         /*
 843          * The MSN might be for a later WQE than the PSN indicates so
 844          * only complete WQEs that the PSN finishes.
 845          */
 846         while (ipath_cmp24(ack_psn, wqe->lpsn) >= 0) {
 847                 /*
 848                  * If this request is a RDMA read or atomic, and the ACK is
 849                  * for a later operation, this ACK NAKs the RDMA read or
 850                  * atomic.  In other words, only a RDMA_READ_LAST or ONLY
 851                  * can ACK a RDMA read and likewise for atomic ops.  Note
 852                  * that the NAK case can only happen if relaxed ordering is
 853                  * used and requests are sent after an RDMA read or atomic
 854                  * is sent but before the response is received.
 855                  */
 856                 if ((wqe->wr.opcode == IB_WR_RDMA_READ &&
 857                      (opcode != OP(RDMA_READ_RESPONSE_LAST) ||
 858                       ipath_cmp24(ack_psn, wqe->lpsn) != 0)) ||
 859                     ((wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
 860                       wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) &&
 861                      (opcode != OP(ATOMIC_ACKNOWLEDGE) ||
 862                       ipath_cmp24(wqe->psn, psn) != 0))) {
 863                         /*
 864                          * The last valid PSN seen is the previous
 865                          * request's.
 866                          */
 867                         update_last_psn(qp, wqe->psn - 1);
 868                         /* Retry this request. */
 869                         ipath_restart_rc(qp, wqe->psn, &wc);
 870                         /*
 871                          * No need to process the ACK/NAK since we are
 872                          * restarting an earlier request.
 873                          */
 874                         goto bail;
 875                 }
 876                 if (qp->s_num_rd_atomic &&
 877                     (wqe->wr.opcode == IB_WR_RDMA_READ ||
 878                      wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
 879                      wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)) {
 880                         qp->s_num_rd_atomic--;
 881                         /* Restart sending task if fence is complete */
 882                         if ((qp->s_flags & IPATH_S_FENCE_PENDING) &&
 883                             !qp->s_num_rd_atomic) {
 884                                 qp->s_flags &= ~IPATH_S_FENCE_PENDING;
 885                                 tasklet_hi_schedule(&qp->s_task);
 886                         } else if (qp->s_flags & IPATH_S_RDMAR_PENDING) {
 887                                 qp->s_flags &= ~IPATH_S_RDMAR_PENDING;
 888                                 tasklet_hi_schedule(&qp->s_task);
 889                         }
 890                 }
 891                 /* Post a send completion queue entry if requested. */
 892                 if (!(qp->s_flags & IPATH_S_SIGNAL_REQ_WR) ||
 893                     (wqe->wr.send_flags & IB_SEND_SIGNALED)) {
 894                         wc.wr_id = wqe->wr.wr_id;
 895                         wc.status = IB_WC_SUCCESS;
 896                         wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
 897                         wc.vendor_err = 0;
 898                         wc.byte_len = wqe->length;
 899                         wc.imm_data = 0;
 900                         wc.qp = &qp->ibqp;
 901                         wc.src_qp = qp->remote_qpn;
 902                         wc.wc_flags = 0;
 903                         wc.pkey_index = 0;
 904                         wc.slid = qp->remote_ah_attr.dlid;
 905                         wc.sl = qp->remote_ah_attr.sl;
 906                         wc.dlid_path_bits = 0;
 907                         wc.port_num = 0;
 908                         ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 0);
 909                 }
 910                 qp->s_retry = qp->s_retry_cnt;
 911                 /*
 912                  * If we are completing a request which is in the process of
 913                  * being resent, we can stop resending it since we know the
 914                  * responder has already seen it.
 915                  */
 916                 if (qp->s_last == qp->s_cur) {
 917                         if (++qp->s_cur >= qp->s_size)
 918                                 qp->s_cur = 0;
 919                         qp->s_last = qp->s_cur;
 920                         if (qp->s_last == qp->s_tail)
 921                                 break;
 922                         wqe = get_swqe_ptr(qp, qp->s_cur);
 923                         qp->s_state = OP(SEND_LAST);
 924                         qp->s_psn = wqe->psn;
 925                 } else {
 926                         if (++qp->s_last >= qp->s_size)
 927                                 qp->s_last = 0;
 928                         if (qp->s_last == qp->s_tail)
 929                                 break;
 930                         wqe = get_swqe_ptr(qp, qp->s_last);
 931                 }
 932         }
 933
 934         switch (aeth >> 29) {
 935         case 0:         /* ACK */
 936                 dev->n_rc_acks++;
 937                 /* If this is a partial ACK, reset the retransmit timer. */
 938                 if (qp->s_last != qp->s_tail) {
 939                         spin_lock(&dev->pending_lock);
 940                         list_add_tail(&qp->timerwait,
 941                                       &dev->pending[dev->pending_index]);
 942                         spin_unlock(&dev->pending_lock);
 943                         /*
 944                          * If we get a partial ACK for a resent operation,
 945                          * we can stop resending the earlier packets and
 946                          * continue with the next packet the receiver wants.
 947                          */
 948                         if (ipath_cmp24(qp->s_psn, psn) <= 0) {
 949                                 reset_psn(qp, psn + 1);
 950                                 tasklet_hi_schedule(&qp->s_task);
 951                         }
 952                 } else if (ipath_cmp24(qp->s_psn, psn) <= 0) {
 953                         qp->s_state = OP(SEND_LAST);
 954                         qp->s_psn = psn + 1;
 955                 }
 956                 ipath_get_credit(qp, aeth);
 957                 qp->s_rnr_retry = qp->s_rnr_retry_cnt;
 958                 qp->s_retry = qp->s_retry_cnt;
 959                 update_last_psn(qp, psn);
 960                 ret = 1;
 961                 goto bail;
 962
 963         case 1:         /* RNR NAK */
 964                 dev->n_rnr_naks++;
 965                 if (qp->s_last == qp->s_tail)
 966                         goto bail;
 967                 if (qp->s_rnr_retry == 0) {
 968                         wc.status = IB_WC_RNR_RETRY_EXC_ERR;
 969                         goto class_b;
 970                 }
 971                 if (qp->s_rnr_retry_cnt < 7)
 972                         qp->s_rnr_retry--;
 973
 974                 /* The last valid PSN is the previous PSN. */
 975                 update_last_psn(qp, psn - 1);
 976
 977                 if (wqe->wr.opcode == IB_WR_RDMA_READ)
 978                         dev->n_rc_resends++;
 979                 else
 980                         dev->n_rc_resends +=
 981                                 (qp->s_psn - psn) & IPATH_PSN_MASK;
 982
 983                 reset_psn(qp, psn);
 984
 985                 qp->s_rnr_timeout =
 986                         ib_ipath_rnr_table[(aeth >> IPATH_AETH_CREDIT_SHIFT) &
 987                                            IPATH_AETH_CREDIT_MASK];
 988                 ipath_insert_rnr_queue(qp);
 989                 goto bail;
 990
 991         case 3:         /* NAK */
 992                 if (qp->s_last == qp->s_tail)
 993                         goto bail;
 994                 /* The last valid PSN is the previous PSN. */
 995                 update_last_psn(qp, psn - 1);
 996                 switch ((aeth >> IPATH_AETH_CREDIT_SHIFT) &
 997                         IPATH_AETH_CREDIT_MASK) {
 998                 case 0: /* PSN sequence error */
 999                         dev->n_seq_naks++;
1000                         /*
1001                          * Back up to the responder's expected PSN.
1002                          * Note that we might get a NAK in the middle of an
1003                          * RDMA READ response which terminates the RDMA
1004                          * READ.
1005                          */
1006                         ipath_restart_rc(qp, psn, &wc);
1007                         break;
1008
1009                 case 1: /* Invalid Request */
1010                         wc.status = IB_WC_REM_INV_REQ_ERR;
1011                         dev->n_other_naks++;
1012                         goto class_b;
1013
1014                 case 2: /* Remote Access Error */
1015                         wc.status = IB_WC_REM_ACCESS_ERR;
1016                         dev->n_other_naks++;
1017                         goto class_b;
1018
1019                 case 3: /* Remote Operation Error */
1020                         wc.status = IB_WC_REM_OP_ERR;
1021                         dev->n_other_naks++;
1022                 class_b:
1023                         wc.wr_id = wqe->wr.wr_id;
1024                         wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
1025                         wc.vendor_err = 0;
1026                         wc.byte_len = 0;
1027                         wc.qp = &qp->ibqp;
1028                         wc.src_qp = qp->remote_qpn;
1029                         wc.pkey_index = 0;
1030                         wc.slid = qp->remote_ah_attr.dlid;
1031                         wc.sl = qp->remote_ah_attr.sl;
1032                         wc.dlid_path_bits = 0;
1033                         wc.port_num = 0;
1034                         ipath_sqerror_qp(qp, &wc);
1035                         break;
1036
1037                 default:
1038                         /* Ignore other reserved NAK error codes */
1039                         goto reserved;
1040                 }
1041                 qp->s_rnr_retry = qp->s_rnr_retry_cnt;
1042                 goto bail;
1043
1044         default:                /* 2: reserved */
1045         reserved:
1046                 /* Ignore reserved NAK codes. */
1047                 goto bail;
1048         }
1049
1050 bail:
1051         return ret;
1052 }
1053
1054 /**
1055  * ipath_rc_rcv_resp - process an incoming RC response packet
1056  * @dev: the device this packet came in on
1057  * @ohdr: the other headers for this packet
1058  * @data: the packet data
1059  * @tlen: the packet length
1060  * @qp: the QP for this packet
1061  * @opcode: the opcode for this packet
1062  * @psn: the packet sequence number for this packet
1063  * @hdrsize: the header length
1064  * @pmtu: the path MTU
1065  * @header_in_data: true if part of the header data is in the data buffer
1066  *
1067  * This is called from ipath_rc_rcv() to process an incoming RC response
1068  * packet for the given QP.
1069  * Called at interrupt level.
1070  */
1071 static inline void ipath_rc_rcv_resp(struct ipath_ibdev *dev,
1072                                      struct ipath_other_headers *ohdr,
1073                                      void *data, u32 tlen,
1074                                      struct ipath_qp *qp,
1075                                      u32 opcode,
1076                                      u32 psn, u32 hdrsize, u32 pmtu,
1077                                      int header_in_data)
1078 {
1079         struct ipath_swqe *wqe;
1080         unsigned long flags;
1081         struct ib_wc wc;
1082         int diff;
1083         u32 pad;
1084         u32 aeth;
1085
1086         spin_lock_irqsave(&qp->s_lock, flags);
1087
1088         /* Ignore invalid responses. */
1089         if (ipath_cmp24(psn, qp->s_next_psn) >= 0)
1090                 goto ack_done;
1091
1092         /* Ignore duplicate responses. */
1093         diff = ipath_cmp24(psn, qp->s_last_psn);
1094         if (unlikely(diff <= 0)) {
1095                 /* Update credits for "ghost" ACKs */
1096                 if (diff == 0 && opcode == OP(ACKNOWLEDGE)) {
1097                         if (!header_in_data)
1098                                 aeth = be32_to_cpu(ohdr->u.aeth);
1099                         else {
1100                                 aeth = be32_to_cpu(((__be32 *) data)[0]);
1101                                 data += sizeof(__be32);
1102                         }
1103                         if ((aeth >> 29) == 0)
1104                                 ipath_get_credit(qp, aeth);
1105                 }
1106                 goto ack_done;
1107         }
1108
1109         if (unlikely(qp->s_last == qp->s_tail))
1110                 goto ack_done;
1111         wqe = get_swqe_ptr(qp, qp->s_last);
1112
1113         switch (opcode) {
1114         case OP(ACKNOWLEDGE):
1115         case OP(ATOMIC_ACKNOWLEDGE):
1116         case OP(RDMA_READ_RESPONSE_FIRST):
1117                 if (!header_in_data)
1118                         aeth = be32_to_cpu(ohdr->u.aeth);
1119                 else {
1120                         aeth = be32_to_cpu(((__be32 *) data)[0]);
1121                         data += sizeof(__be32);
1122                 }
1123                 if (opcode == OP(ATOMIC_ACKNOWLEDGE)) {
1124                         u64 val;
1125
1126                         if (!header_in_data) {
1127                                 __be32 *p = ohdr->u.at.atomic_ack_eth;
1128
1129                                 val = ((u64) be32_to_cpu(p[0]) << 32) |
1130                                         be32_to_cpu(p[1]);
1131                         } else
1132                                 val = be64_to_cpu(((__be64 *) data)[0]);
1133                         *(u64 *) wqe->sg_list[0].vaddr = val;
1134                 }
1135                 if (!do_rc_ack(qp, aeth, psn, opcode) ||
1136                     opcode != OP(RDMA_READ_RESPONSE_FIRST))
1137                         goto ack_done;
1138                 hdrsize += 4;
1139                 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1140                         goto ack_op_err;
1141                 /*
1142                  * If this is a response to a resent RDMA read, we
1143                  * have to be careful to copy the data to the right
1144                  * location.
1145                  */
1146                 qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
1147                                                   wqe, psn, pmtu);
1148                 goto read_middle;
1149
1150         case OP(RDMA_READ_RESPONSE_MIDDLE):
1151                 /* no AETH, no ACK */
1152                 if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) {
1153                         dev->n_rdma_seq++;
1154                         ipath_restart_rc(qp, qp->s_last_psn + 1, &wc);
1155                         goto ack_done;
1156                 }
1157                 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1158                         goto ack_op_err;
1159         read_middle:
1160                 if (unlikely(tlen != (hdrsize + pmtu + 4)))
1161                         goto ack_len_err;
1162                 if (unlikely(pmtu >= qp->s_rdma_read_len))
1163                         goto ack_len_err;
1164
1165                 /* We got a response so update the timeout. */
1166                 spin_lock(&dev->pending_lock);
1167                 if (qp->s_rnr_timeout == 0 && !list_empty(&qp->timerwait))
1168                         list_move_tail(&qp->timerwait,
1169                                        &dev->pending[dev->pending_index]);
1170                 spin_unlock(&dev->pending_lock);
1171                 /*
1172                  * Update the RDMA receive state but do the copy w/o
1173                  * holding the locks and blocking interrupts.
1174                  */
1175                 qp->s_rdma_read_len -= pmtu;
1176                 update_last_psn(qp, psn);
1177                 spin_unlock_irqrestore(&qp->s_lock, flags);
1178                 ipath_copy_sge(&qp->s_rdma_read_sge, data, pmtu);
1179                 goto bail;
1180
1181         case OP(RDMA_READ_RESPONSE_ONLY):
1182                 if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) {
1183                         dev->n_rdma_seq++;
1184                         ipath_restart_rc(qp, qp->s_last_psn + 1, &wc);
1185                         goto ack_done;
1186                 }
1187                 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1188                         goto ack_op_err;
1189                 /* Get the number of bytes the message was padded by. */
1190                 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1191                 /*
1192                  * Check that the data size is >= 0 && <= pmtu.
1193                  * Remember to account for the AETH header (4) and
1194                  * ICRC (4).
1195                  */
1196                 if (unlikely(tlen < (hdrsize + pad + 8)))
1197                         goto ack_len_err;
1198                 /*
1199                  * If this is a response to a resent RDMA read, we
1200                  * have to be careful to copy the data to the right
1201                  * location.
1202                  */
1203                 qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
1204                                                   wqe, psn, pmtu);
1205                 goto read_last;
1206
1207         case OP(RDMA_READ_RESPONSE_LAST):
1208                 /* ACKs READ req. */
1209                 if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) {
1210                         dev->n_rdma_seq++;
1211                         ipath_restart_rc(qp, qp->s_last_psn + 1, &wc);
1212                         goto ack_done;
1213                 }
1214                 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1215                         goto ack_op_err;
1216                 /* Get the number of bytes the message was padded by. */
1217                 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1218                 /*
1219                  * Check that the data size is >= 1 && <= pmtu.
1220                  * Remember to account for the AETH header (4) and
1221                  * ICRC (4).
1222                  */
1223                 if (unlikely(tlen <= (hdrsize + pad + 8)))
1224                         goto ack_len_err;
1225         read_last:
1226                 tlen -= hdrsize + pad + 8;
1227                 if (unlikely(tlen != qp->s_rdma_read_len))
1228                         goto ack_len_err;
1229                 if (!header_in_data)
1230                         aeth = be32_to_cpu(ohdr->u.aeth);
1231                 else {
1232                         aeth = be32_to_cpu(((__be32 *) data)[0]);
1233                         data += sizeof(__be32);
1234                 }
1235                 ipath_copy_sge(&qp->s_rdma_read_sge, data, tlen);
1236                 (void) do_rc_ack(qp, aeth, psn, OP(RDMA_READ_RESPONSE_LAST));
1237                 goto ack_done;
1238         }
1239
1240 ack_done:
1241         spin_unlock_irqrestore(&qp->s_lock, flags);
1242         goto bail;
1243
1244 ack_op_err:
1245         wc.status = IB_WC_LOC_QP_OP_ERR;
1246         goto ack_err;
1247
1248 ack_len_err:
1249         wc.status = IB_WC_LOC_LEN_ERR;
1250 ack_err:
1251         wc.wr_id = wqe->wr.wr_id;
1252         wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
1253         wc.vendor_err = 0;
1254         wc.byte_len = 0;
1255         wc.imm_data = 0;
1256         wc.qp = &qp->ibqp;
1257         wc.src_qp = qp->remote_qpn;
1258         wc.wc_flags = 0;
1259         wc.pkey_index = 0;
1260         wc.slid = qp->remote_ah_attr.dlid;
1261         wc.sl = qp->remote_ah_attr.sl;
1262         wc.dlid_path_bits = 0;
1263         wc.port_num = 0;
1264         ipath_sqerror_qp(qp, &wc);
1265 bail:
1266         return;
1267 }
1268
1269 /**
1270  * ipath_rc_rcv_error - process an incoming duplicate or error RC packet
1271  * @dev: the device this packet came in on
1272  * @ohdr: the other headers for this packet
1273  * @data: the packet data
1274  * @qp: the QP for this packet
1275  * @opcode: the opcode for this packet
1276  * @psn: the packet sequence number for this packet
1277  * @diff: the difference between the PSN and the expected PSN
1278  * @header_in_data: true if part of the header data is in the data buffer
1279  *
1280  * This is called from ipath_rc_rcv() to process an unexpected
1281  * incoming RC packet for the given QP.
1282  * Called at interrupt level.
1283  * Return 1 if no more processing is needed; otherwise return 0 to
1284  * schedule a response to be sent.
1285  */
1286 static inline int ipath_rc_rcv_error(struct ipath_ibdev *dev,
1287                                      struct ipath_other_headers *ohdr,
1288                                      void *data,
1289                                      struct ipath_qp *qp,
1290                                      u32 opcode,
1291                                      u32 psn,
1292                                      int diff,
1293                                      int header_in_data)
1294 {
1295         struct ipath_ack_entry *e;
1296         u8 i, prev;
1297         int old_req;
1298         unsigned long flags;
1299
1300         if (diff > 0) {
1301                 /*
1302                  * Packet sequence error.
1303                  * A NAK will ACK earlier sends and RDMA writes.
1304                  * Don't queue the NAK if we already sent one.
1305                  */
1306                 if (!qp->r_nak_state) {
1307                         qp->r_nak_state = IB_NAK_PSN_ERROR;
1308                         /* Use the expected PSN. */
1309                         qp->r_ack_psn = qp->r_psn;
1310                         goto send_ack;
1311                 }
1312                 goto done;
1313         }
1314
1315         /*
1316          * Handle a duplicate request.  Don't re-execute SEND, RDMA
1317          * write or atomic op.  Don't NAK errors, just silently drop
1318          * the duplicate request.  Note that r_sge, r_len, and
1319          * r_rcv_len may be in use so don't modify them.
1320          *
1321          * We are supposed to ACK the earliest duplicate PSN but we
1322          * can coalesce an outstanding duplicate ACK.  We have to
1323          * send the earliest so that RDMA reads can be restarted at
1324          * the requester's expected PSN.
1325          *
1326          * First, find where this duplicate PSN falls within the
1327          * ACKs previously sent.
1328          */
1329         psn &= IPATH_PSN_MASK;
1330         e = NULL;
1331         old_req = 1;
1332         spin_lock_irqsave(&qp->s_lock, flags);
1333         for (i = qp->r_head_ack_queue; ; i = prev) {
1334                 if (i == qp->s_tail_ack_queue)
1335                         old_req = 0;
1336                 if (i)
1337                         prev = i - 1;
1338                 else
1339                         prev = IPATH_MAX_RDMA_ATOMIC;
1340                 if (prev == qp->r_head_ack_queue) {
1341                         e = NULL;
1342                         break;
1343                 }
1344                 e = &qp->s_ack_queue[prev];
1345                 if (!e->opcode) {
1346                         e = NULL;
1347                         break;
1348                 }
1349                 if (ipath_cmp24(psn, e->psn) >= 0)
1350                         break;
1351         }
1352         switch (opcode) {
1353         case OP(RDMA_READ_REQUEST): {
1354                 struct ib_reth *reth;
1355                 u32 offset;
1356                 u32 len;
1357
1358                 /*
1359                  * If we didn't find the RDMA read request in the ack queue,
1360                  * or the send tasklet is already backed up to send an
1361                  * earlier entry, we can ignore this request.
1362                  */
1363                 if (!e || e->opcode != OP(RDMA_READ_REQUEST) || old_req)
1364                         goto unlock_done;
1365                 /* RETH comes after BTH */
1366                 if (!header_in_data)
1367                         reth = &ohdr->u.rc.reth;
1368                 else {
1369                         reth = (struct ib_reth *)data;
1370                         data += sizeof(*reth);
1371                 }
1372                 /*
1373                  * Address range must be a subset of the original
1374                  * request and start on pmtu boundaries.
1375                  * We reuse the old ack_queue slot since the requester
1376                  * should not back up and request an earlier PSN for the
1377                  * same request.
1378                  */
1379                 offset = ((psn - e->psn) & IPATH_PSN_MASK) *
1380                         ib_mtu_enum_to_int(qp->path_mtu);
1381                 len = be32_to_cpu(reth->length);
1382                 if (unlikely(offset + len > e->rdma_sge.sge.sge_length))
1383                         goto unlock_done;
1384                 if (len != 0) {
1385                         u32 rkey = be32_to_cpu(reth->rkey);
1386                         u64 vaddr = be64_to_cpu(reth->vaddr);
1387                         int ok;
1388
1389                         ok = ipath_rkey_ok(qp, &e->rdma_sge,
1390                                            len, vaddr, rkey,
1391                                            IB_ACCESS_REMOTE_READ);
1392                         if (unlikely(!ok))
1393                                 goto unlock_done;
1394                 } else {
1395                         e->rdma_sge.sg_list = NULL;
1396                         e->rdma_sge.num_sge = 0;
1397                         e->rdma_sge.sge.mr = NULL;
1398                         e->rdma_sge.sge.vaddr = NULL;
1399                         e->rdma_sge.sge.length = 0;
1400                         e->rdma_sge.sge.sge_length = 0;
1401                 }
1402                 e->psn = psn;
1403                 qp->s_ack_state = OP(ACKNOWLEDGE);
1404                 qp->s_tail_ack_queue = prev;
1405                 break;
1406         }
1407
1408         case OP(COMPARE_SWAP):
1409         case OP(FETCH_ADD): {
1410                 /*
1411                  * If we didn't find the atomic request in the ack queue
1412                  * or the send tasklet is already backed up to send an
1413                  * earlier entry, we can ignore this request.
1414                  */
1415                 if (!e || e->opcode != (u8) opcode || old_req)
1416                         goto unlock_done;
1417                 qp->s_ack_state = OP(ACKNOWLEDGE);
1418                 qp->s_tail_ack_queue = prev;
1419                 break;
1420         }
1421
1422         default:
1423                 if (old_req)
1424                         goto unlock_done;
1425                 /*
1426                  * Resend the most recent ACK if this request is
1427                  * after all the previous RDMA reads and atomics.
1428                  */
1429                 if (i == qp->r_head_ack_queue) {
1430                         spin_unlock_irqrestore(&qp->s_lock, flags);
1431                         qp->r_nak_state = 0;
1432                         qp->r_ack_psn = qp->r_psn - 1;
1433                         goto send_ack;
1434                 }
1435                 /*
1436                  * Resend the RDMA read or atomic op which
1437                  * ACKs this duplicate request.
1438                  */
1439                 qp->s_ack_state = OP(ACKNOWLEDGE);
1440                 qp->s_tail_ack_queue = i;
1441                 break;
1442         }
1443         qp->r_nak_state = 0;
1444         spin_unlock_irq(&qp->s_lock);
1445         tasklet_hi_schedule(&qp->s_task);
1446
1447 unlock_done:
1448         spin_unlock_irqrestore(&qp->s_lock, flags);
1449 done:
1450         return 1;
1451
1452 send_ack:
1453         return 0;
1454 }
1455
1456 static void ipath_rc_error(struct ipath_qp *qp, enum ib_wc_status err)
1457 {
1458         unsigned long flags;
1459
1460         spin_lock_irqsave(&qp->s_lock, flags);
1461         qp->state = IB_QPS_ERR;
1462         ipath_error_qp(qp, err);
1463         spin_unlock_irqrestore(&qp->s_lock, flags);
1464 }
1465
1466 /**
1467  * ipath_rc_rcv - process an incoming RC packet
1468  * @dev: the device this packet came in on
1469  * @hdr: the header of this packet
1470  * @has_grh: true if the header has a GRH
1471  * @data: the packet data
1472  * @tlen: the packet length
1473  * @qp: the QP for this packet
1474  *
1475  * This is called from ipath_qp_rcv() to process an incoming RC packet
1476  * for the given QP.
1477  * Called at interrupt level.
1478  */
1479 void ipath_rc_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
1480                   int has_grh, void *data, u32 tlen, struct ipath_qp *qp)
1481 {
1482         struct ipath_other_headers *ohdr;
1483         u32 opcode;
1484         u32 hdrsize;
1485         u32 psn;
1486         u32 pad;
1487         struct ib_wc wc;
1488         u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
1489         int diff;
1490         struct ib_reth *reth;
1491         int header_in_data;
1492
1493         /* Validate the SLID. See Ch. 9.6.1.5 */
1494         if (unlikely(be16_to_cpu(hdr->lrh[3]) != qp->remote_ah_attr.dlid))
1495                 goto done;
1496
1497         /* Check for GRH */
1498         if (!has_grh) {
1499                 ohdr = &hdr->u.oth;
1500                 hdrsize = 8 + 12;       /* LRH + BTH */
1501                 psn = be32_to_cpu(ohdr->bth[2]);
1502                 header_in_data = 0;
1503         } else {
1504                 ohdr = &hdr->u.l.oth;
1505                 hdrsize = 8 + 40 + 12;  /* LRH + GRH + BTH */
1506                 /*
1507                  * The header with GRH is 60 bytes and the core driver sets
1508                  * the eager header buffer size to 56 bytes so the last 4
1509                  * bytes of the BTH header (PSN) is in the data buffer.
1510                  */
1511                 header_in_data = dev->dd->ipath_rcvhdrentsize == 16;
1512                 if (header_in_data) {
1513                         psn = be32_to_cpu(((__be32 *) data)[0]);
1514                         data += sizeof(__be32);
1515                 } else
1516                         psn = be32_to_cpu(ohdr->bth[2]);
1517         }
1518
1519         /*
1520          * Process responses (ACKs) before anything else.  Note that the
1521          * packet sequence number will be for something in the send work
1522          * queue rather than the expected receive packet sequence number.
1523          * In other words, this QP is the requester.
1524          */
1525         opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
1526         if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
1527             opcode <= OP(ATOMIC_ACKNOWLEDGE)) {
1528                 ipath_rc_rcv_resp(dev, ohdr, data, tlen, qp, opcode, psn,
1529                                   hdrsize, pmtu, header_in_data);
1530                 goto done;
1531         }
1532
1533         /* Compute 24 bits worth of difference. */
1534         diff = ipath_cmp24(psn, qp->r_psn);
1535         if (unlikely(diff)) {
1536                 if (ipath_rc_rcv_error(dev, ohdr, data, qp, opcode,
1537                                        psn, diff, header_in_data))
1538                         goto done;
1539                 goto send_ack;
1540         }
1541
1542         /* Check for opcode sequence errors. */
1543         switch (qp->r_state) {
1544         case OP(SEND_FIRST):
1545         case OP(SEND_MIDDLE):
1546                 if (opcode == OP(SEND_MIDDLE) ||
1547                     opcode == OP(SEND_LAST) ||
1548                     opcode == OP(SEND_LAST_WITH_IMMEDIATE))
1549                         break;
1550         nack_inv:
1551                 ipath_rc_error(qp, IB_WC_REM_INV_REQ_ERR);
1552                 qp->r_nak_state = IB_NAK_INVALID_REQUEST;
1553                 qp->r_ack_psn = qp->r_psn;
1554                 goto send_ack;
1555
1556         case OP(RDMA_WRITE_FIRST):
1557         case OP(RDMA_WRITE_MIDDLE):
1558                 if (opcode == OP(RDMA_WRITE_MIDDLE) ||
1559                     opcode == OP(RDMA_WRITE_LAST) ||
1560                     opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
1561                         break;
1562                 goto nack_inv;
1563
1564         default:
1565                 if (opcode == OP(SEND_MIDDLE) ||
1566                     opcode == OP(SEND_LAST) ||
1567                     opcode == OP(SEND_LAST_WITH_IMMEDIATE) ||
1568                     opcode == OP(RDMA_WRITE_MIDDLE) ||
1569                     opcode == OP(RDMA_WRITE_LAST) ||
1570                     opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
1571                         goto nack_inv;
1572                 /*
1573                  * Note that it is up to the requester to not send a new
1574                  * RDMA read or atomic operation before receiving an ACK
1575                  * for the previous operation.
1576                  */
1577                 break;
1578         }
1579
1580         wc.imm_data = 0;
1581         wc.wc_flags = 0;
1582
1583         /* OK, process the packet. */
1584         switch (opcode) {
1585         case OP(SEND_FIRST):
1586                 if (!ipath_get_rwqe(qp, 0)) {
1587                 rnr_nak:
1588                         /*
1589                          * A RNR NAK will ACK earlier sends and RDMA writes.
1590                          * Don't queue the NAK if a RDMA read or atomic
1591                          * is pending though.
1592                          */
1593                         if (qp->r_nak_state)
1594                                 goto done;
1595                         qp->r_nak_state = IB_RNR_NAK | qp->r_min_rnr_timer;
1596                         qp->r_ack_psn = qp->r_psn;
1597                         goto send_ack;
1598                 }
1599                 qp->r_rcv_len = 0;
1600                 /* FALLTHROUGH */
1601         case OP(SEND_MIDDLE):
1602         case OP(RDMA_WRITE_MIDDLE):
1603         send_middle:
1604                 /* Check for invalid length PMTU or posted rwqe len. */
1605                 if (unlikely(tlen != (hdrsize + pmtu + 4)))
1606                         goto nack_inv;
1607                 qp->r_rcv_len += pmtu;
1608                 if (unlikely(qp->r_rcv_len > qp->r_len))
1609                         goto nack_inv;
1610                 ipath_copy_sge(&qp->r_sge, data, pmtu);
1611                 break;
1612
1613         case OP(RDMA_WRITE_LAST_WITH_IMMEDIATE):
1614                 /* consume RWQE */
1615                 if (!ipath_get_rwqe(qp, 1))
1616                         goto rnr_nak;
1617                 goto send_last_imm;
1618
1619         case OP(SEND_ONLY):
1620         case OP(SEND_ONLY_WITH_IMMEDIATE):
1621                 if (!ipath_get_rwqe(qp, 0))
1622                         goto rnr_nak;
1623                 qp->r_rcv_len = 0;
1624                 if (opcode == OP(SEND_ONLY))
1625                         goto send_last;
1626                 /* FALLTHROUGH */
1627         case OP(SEND_LAST_WITH_IMMEDIATE):
1628         send_last_imm:
1629                 if (header_in_data) {
1630                         wc.imm_data = *(__be32 *) data;
1631                         data += sizeof(__be32);
1632                 } else {
1633                         /* Immediate data comes after BTH */
1634                         wc.imm_data = ohdr->u.imm_data;
1635                 }
1636                 hdrsize += 4;
1637                 wc.wc_flags = IB_WC_WITH_IMM;
1638                 /* FALLTHROUGH */
1639         case OP(SEND_LAST):
1640         case OP(RDMA_WRITE_LAST):
1641         send_last:
1642                 /* Get the number of bytes the message was padded by. */
1643                 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1644                 /* Check for invalid length. */
1645                 /* XXX LAST len should be >= 1 */
1646                 if (unlikely(tlen < (hdrsize + pad + 4)))
1647                         goto nack_inv;
1648                 /* Don't count the CRC. */
1649                 tlen -= (hdrsize + pad + 4);
1650                 wc.byte_len = tlen + qp->r_rcv_len;
1651                 if (unlikely(wc.byte_len > qp->r_len))
1652                         goto nack_inv;
1653                 ipath_copy_sge(&qp->r_sge, data, tlen);
1654                 qp->r_msn++;
1655                 if (!qp->r_wrid_valid)
1656                         break;
1657                 qp->r_wrid_valid = 0;
1658                 wc.wr_id = qp->r_wr_id;
1659                 wc.status = IB_WC_SUCCESS;
1660                 wc.opcode = IB_WC_RECV;
1661                 wc.vendor_err = 0;
1662                 wc.qp = &qp->ibqp;
1663                 wc.src_qp = qp->remote_qpn;
1664                 wc.pkey_index = 0;
1665                 wc.slid = qp->remote_ah_attr.dlid;
1666                 wc.sl = qp->remote_ah_attr.sl;
1667                 wc.dlid_path_bits = 0;
1668                 wc.port_num = 0;
1669                 /* Signal completion event if the solicited bit is set. */
1670                 ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
1671                                (ohdr->bth[0] &
1672                                 __constant_cpu_to_be32(1 << 23)) != 0);
1673                 break;
1674
1675         case OP(RDMA_WRITE_FIRST):
1676         case OP(RDMA_WRITE_ONLY):
1677         case OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE):
1678                 /* consume RWQE */
1679                 /* RETH comes after BTH */
1680                 if (!header_in_data)
1681                         reth = &ohdr->u.rc.reth;
1682                 else {
1683                         reth = (struct ib_reth *)data;
1684                         data += sizeof(*reth);
1685                 }
1686                 hdrsize += sizeof(*reth);
1687                 qp->r_len = be32_to_cpu(reth->length);
1688                 qp->r_rcv_len = 0;
1689                 if (qp->r_len != 0) {
1690                         u32 rkey = be32_to_cpu(reth->rkey);
1691                         u64 vaddr = be64_to_cpu(reth->vaddr);
1692                         int ok;
1693
1694                         /* Check rkey & NAK */
1695                         ok = ipath_rkey_ok(qp, &qp->r_sge,
1696                                            qp->r_len, vaddr, rkey,
1697                                            IB_ACCESS_REMOTE_WRITE);
1698                         if (unlikely(!ok))
1699                                 goto nack_acc;
1700                 } else {
1701                         qp->r_sge.sg_list = NULL;
1702                         qp->r_sge.sge.mr = NULL;
1703                         qp->r_sge.sge.vaddr = NULL;
1704                         qp->r_sge.sge.length = 0;
1705                         qp->r_sge.sge.sge_length = 0;
1706                 }
1707                 if (unlikely(!(qp->qp_access_flags &
1708                                IB_ACCESS_REMOTE_WRITE)))
1709                         goto nack_acc;
1710                 if (opcode == OP(RDMA_WRITE_FIRST))
1711                         goto send_middle;
1712                 else if (opcode == OP(RDMA_WRITE_ONLY))
1713                         goto send_last;
1714                 if (!ipath_get_rwqe(qp, 1))
1715                         goto rnr_nak;
1716                 goto send_last_imm;
1717
1718         case OP(RDMA_READ_REQUEST): {
1719                 struct ipath_ack_entry *e;
1720                 u32 len;
1721                 u8 next;
1722
1723                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ)))
1724                         goto nack_acc;
1725                 next = qp->r_head_ack_queue + 1;
1726                 if (next > IPATH_MAX_RDMA_ATOMIC)
1727                         next = 0;
1728                 if (unlikely(next == qp->s_tail_ack_queue))
1729                         goto nack_inv;
1730                 e = &qp->s_ack_queue[qp->r_head_ack_queue];
1731                 /* RETH comes after BTH */
1732                 if (!header_in_data)
1733                         reth = &ohdr->u.rc.reth;
1734                 else {
1735                         reth = (struct ib_reth *)data;
1736                         data += sizeof(*reth);
1737                 }
1738                 len = be32_to_cpu(reth->length);
1739                 if (len) {
1740                         u32 rkey = be32_to_cpu(reth->rkey);
1741                         u64 vaddr = be64_to_cpu(reth->vaddr);
1742                         int ok;
1743
1744                         /* Check rkey & NAK */
1745                         ok = ipath_rkey_ok(qp, &e->rdma_sge, len, vaddr,
1746                                            rkey, IB_ACCESS_REMOTE_READ);
1747                         if (unlikely(!ok))
1748                                 goto nack_acc;
1749                         /*
1750                          * Update the next expected PSN.  We add 1 later
1751                          * below, so only add the remainder here.
1752                          */
1753                         if (len > pmtu)
1754                                 qp->r_psn += (len - 1) / pmtu;
1755                 } else {
1756                         e->rdma_sge.sg_list = NULL;
1757                         e->rdma_sge.num_sge = 0;
1758                         e->rdma_sge.sge.mr = NULL;
1759                         e->rdma_sge.sge.vaddr = NULL;
1760                         e->rdma_sge.sge.length = 0;
1761                         e->rdma_sge.sge.sge_length = 0;
1762                 }
1763                 e->opcode = opcode;
1764                 e->psn = psn;
1765                 /*
1766                  * We need to increment the MSN here instead of when we
1767                  * finish sending the result since a duplicate request would
1768                  * increment it more than once.
1769                  */
1770                 qp->r_msn++;
1771                 qp->r_psn++;
1772                 qp->r_state = opcode;
1773                 qp->r_nak_state = 0;
1774                 barrier();
1775                 qp->r_head_ack_queue = next;
1776
1777                 /* Call ipath_do_rc_send() in another thread. */
1778                 tasklet_hi_schedule(&qp->s_task);
1779
1780                 goto done;
1781         }
1782
1783         case OP(COMPARE_SWAP):
1784         case OP(FETCH_ADD): {
1785                 struct ib_atomic_eth *ateth;
1786                 struct ipath_ack_entry *e;
1787                 u64 vaddr;
1788                 atomic64_t *maddr;
1789                 u64 sdata;
1790                 u32 rkey;
1791                 u8 next;
1792
1793                 if (unlikely(!(qp->qp_access_flags &
1794                                IB_ACCESS_REMOTE_ATOMIC)))
1795                         goto nack_acc;
1796                 next = qp->r_head_ack_queue + 1;
1797                 if (next > IPATH_MAX_RDMA_ATOMIC)
1798                         next = 0;
1799                 if (unlikely(next == qp->s_tail_ack_queue))
1800                         goto nack_inv;
1801                 if (!header_in_data)
1802                         ateth = &ohdr->u.atomic_eth;
1803                 else
1804                         ateth = (struct ib_atomic_eth *)data;
1805                 vaddr = ((u64) be32_to_cpu(ateth->vaddr[0]) << 32) |
1806                         be32_to_cpu(ateth->vaddr[1]);
1807                 if (unlikely(vaddr & (sizeof(u64) - 1)))
1808                         goto nack_inv;
1809                 rkey = be32_to_cpu(ateth->rkey);
1810                 /* Check rkey & NAK */
1811                 if (unlikely(!ipath_rkey_ok(qp, &qp->r_sge,
1812                                             sizeof(u64), vaddr, rkey,
1813                                             IB_ACCESS_REMOTE_ATOMIC)))
1814                         goto nack_acc;
1815                 /* Perform atomic OP and save result. */
1816                 maddr = (atomic64_t *) qp->r_sge.sge.vaddr;
1817                 sdata = be64_to_cpu(ateth->swap_data);
1818                 e = &qp->s_ack_queue[qp->r_head_ack_queue];
1819                 e->atomic_data = (opcode == OP(FETCH_ADD)) ?
1820                         (u64) atomic64_add_return(sdata, maddr) - sdata :
1821                         (u64) cmpxchg((u64 *) qp->r_sge.sge.vaddr,
1822                                       be64_to_cpu(ateth->compare_data),
1823                                       sdata);
1824                 e->opcode = opcode;
1825                 e->psn = psn & IPATH_PSN_MASK;
1826                 qp->r_msn++;
1827                 qp->r_psn++;
1828                 qp->r_state = opcode;
1829                 qp->r_nak_state = 0;
1830                 barrier();
1831                 qp->r_head_ack_queue = next;
1832
1833                 /* Call ipath_do_rc_send() in another thread. */
1834                 tasklet_hi_schedule(&qp->s_task);
1835
1836                 goto done;
1837         }
1838
1839         default:
1840                 /* NAK unknown opcodes. */
1841                 goto nack_inv;
1842         }
1843         qp->r_psn++;
1844         qp->r_state = opcode;
1845         qp->r_ack_psn = psn;
1846         qp->r_nak_state = 0;
1847         /* Send an ACK if requested or required. */
1848         if (psn & (1 << 31))
1849                 goto send_ack;
1850         goto done;
1851
1852 nack_acc:
1853         ipath_rc_error(qp, IB_WC_REM_ACCESS_ERR);
1854         qp->r_nak_state = IB_NAK_REMOTE_ACCESS_ERROR;
1855         qp->r_ack_psn = qp->r_psn;
1856
1857 send_ack:
1858         send_rc_ack(qp);
1859
1860 done:
1861         return;
1862 }