drivers/infiniband/hw/ipath/ipath_rc.c

   1 /*
   2  * Copyright (c) 2006, 2007 QLogic Corporation. 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_ibdev *dev, struct ipath_qp *qp,
  85                              struct ipath_other_headers *ohdr, u32 pmtu)
  86 {
  87         struct ipath_ack_entry *e;
  88         u32 hwords;
  89         u32 len;
  90         u32 bth0;
  91         u32 bth2;
  92
  93         /* header size in 32-bit words LRH+BTH = (8+12)/4. */
  94         hwords = 5;
  95
  96         switch (qp->s_ack_state) {
  97         case OP(RDMA_READ_RESPONSE_LAST):
  98         case OP(RDMA_READ_RESPONSE_ONLY):
  99         case OP(ATOMIC_ACKNOWLEDGE):
 100                 /*
 101                  * We can increment the tail pointer now that the last
 102                  * response has been sent instead of only being
 103                  * constructed.
 104                  */
 105                 if (++qp->s_tail_ack_queue > IPATH_MAX_RDMA_ATOMIC)
 106                         qp->s_tail_ack_queue = 0;
 107                 /* FALLTHROUGH */
 108         case OP(SEND_ONLY):
 109         case OP(ACKNOWLEDGE):
 110                 /* Check for no next entry in the queue. */
 111                 if (qp->r_head_ack_queue == qp->s_tail_ack_queue) {
 112                         if (qp->s_flags & IPATH_S_ACK_PENDING)
 113                                 goto normal;
 114                         qp->s_ack_state = OP(ACKNOWLEDGE);
 115                         goto bail;
 116                 }
 117
 118                 e = &qp->s_ack_queue[qp->s_tail_ack_queue];
 119                 if (e->opcode == OP(RDMA_READ_REQUEST)) {
 120                         /* Copy SGE state in case we need to resend */
 121                         qp->s_ack_rdma_sge = e->rdma_sge;
 122                         qp->s_cur_sge = &qp->s_ack_rdma_sge;
 123                         len = e->rdma_sge.sge.sge_length;
 124                         if (len > pmtu) {
 125                                 len = pmtu;
 126                                 qp->s_ack_state = OP(RDMA_READ_RESPONSE_FIRST);
 127                         } else {
 128                                 qp->s_ack_state = OP(RDMA_READ_RESPONSE_ONLY);
 129                                 e->sent = 1;
 130                         }
 131                         ohdr->u.aeth = ipath_compute_aeth(qp);
 132                         hwords++;
 133                         qp->s_ack_rdma_psn = e->psn;
 134                         bth2 = qp->s_ack_rdma_psn++ & IPATH_PSN_MASK;
 135                 } else {
 136                         /* COMPARE_SWAP or FETCH_ADD */
 137                         qp->s_cur_sge = NULL;
 138                         len = 0;
 139                         qp->s_ack_state = OP(ATOMIC_ACKNOWLEDGE);
 140                         ohdr->u.at.aeth = ipath_compute_aeth(qp);
 141                         ohdr->u.at.atomic_ack_eth[0] =
 142                                 cpu_to_be32(e->atomic_data >> 32);
 143                         ohdr->u.at.atomic_ack_eth[1] =
 144                                 cpu_to_be32(e->atomic_data);
 145                         hwords += sizeof(ohdr->u.at) / sizeof(u32);
 146                         bth2 = e->psn;
 147                         e->sent = 1;
 148                 }
 149                 bth0 = qp->s_ack_state << 24;
 150                 break;
 151
 152         case OP(RDMA_READ_RESPONSE_FIRST):
 153                 qp->s_ack_state = OP(RDMA_READ_RESPONSE_MIDDLE);
 154                 /* FALLTHROUGH */
 155         case OP(RDMA_READ_RESPONSE_MIDDLE):
 156                 len = qp->s_ack_rdma_sge.sge.sge_length;
 157                 if (len > pmtu)
 158                         len = pmtu;
 159                 else {
 160                         ohdr->u.aeth = ipath_compute_aeth(qp);
 161                         hwords++;
 162                         qp->s_ack_state = OP(RDMA_READ_RESPONSE_LAST);
 163                         qp->s_ack_queue[qp->s_tail_ack_queue].sent = 1;
 164                 }
 165                 bth0 = qp->s_ack_state << 24;
 166                 bth2 = qp->s_ack_rdma_psn++ & IPATH_PSN_MASK;
 167                 break;
 168
 169         default:
 170         normal:
 171                 /*
 172                  * Send a regular ACK.
 173                  * Set the s_ack_state so we wait until after sending
 174                  * the ACK before setting s_ack_state to ACKNOWLEDGE
 175                  * (see above).
 176                  */
 177                 qp->s_ack_state = OP(SEND_ONLY);
 178                 qp->s_flags &= ~IPATH_S_ACK_PENDING;
 179                 qp->s_cur_sge = NULL;
 180                 if (qp->s_nak_state)
 181                         ohdr->u.aeth =
 182                                 cpu_to_be32((qp->r_msn & IPATH_MSN_MASK) |
 183                                             (qp->s_nak_state <<
 184                                              IPATH_AETH_CREDIT_SHIFT));
 185                 else
 186                         ohdr->u.aeth = ipath_compute_aeth(qp);
 187                 hwords++;
 188                 len = 0;
 189                 bth0 = OP(ACKNOWLEDGE) << 24;
 190                 bth2 = qp->s_ack_psn & IPATH_PSN_MASK;
 191         }
 192         qp->s_hdrwords = hwords;
 193         qp->s_cur_size = len;
 194         ipath_make_ruc_header(dev, qp, ohdr, bth0, bth2);
 195         return 1;
 196
 197 bail:
 198         return 0;
 199 }
 200
 201 /**
 202  * ipath_make_rc_req - construct a request packet (SEND, RDMA r/w, ATOMIC)
 203  * @qp: a pointer to the QP
 204  *
 205  * Return 1 if constructed; otherwise, return 0.
 206  */
 207 int ipath_make_rc_req(struct ipath_qp *qp)
 208 {
 209         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
 210         struct ipath_other_headers *ohdr;
 211         struct ipath_sge_state *ss;
 212         struct ipath_swqe *wqe;
 213         u32 hwords;
 214         u32 len;
 215         u32 bth0;
 216         u32 bth2;
 217         u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
 218         char newreq;
 219         unsigned long flags;
 220         int ret = 0;
 221
 222         ohdr = &qp->s_hdr.u.oth;
 223         if (qp->remote_ah_attr.ah_flags & IB_AH_GRH)
 224                 ohdr = &qp->s_hdr.u.l.oth;
 225
 226         /*
 227          * The lock is needed to synchronize between the sending tasklet,
 228          * the receive interrupt handler, and timeout resends.
 229          */
 230         spin_lock_irqsave(&qp->s_lock, flags);
 231
 232         /* Sending responses has higher priority over sending requests. */
 233         if ((qp->r_head_ack_queue != qp->s_tail_ack_queue ||
 234              (qp->s_flags & IPATH_S_ACK_PENDING) ||
 235              qp->s_ack_state != OP(ACKNOWLEDGE)) &&
 236             ipath_make_rc_ack(dev, qp, ohdr, pmtu))
 237                 goto done;
 238
 239         if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK) ||
 240             qp->s_rnr_timeout || qp->s_wait_credit)
 241                 goto bail;
 242
 243         /* Limit the number of packets sent without an ACK. */
 244         if (ipath_cmp24(qp->s_psn, qp->s_last_psn + IPATH_PSN_CREDIT) > 0) {
 245                 qp->s_wait_credit = 1;
 246                 dev->n_rc_stalls++;
 247                 goto bail;
 248         }
 249
 250         /* header size in 32-bit words LRH+BTH = (8+12)/4. */
 251         hwords = 5;
 252         bth0 = 1 << 22; /* Set M bit */
 253
 254         /* Send a request. */
 255         wqe = get_swqe_ptr(qp, qp->s_cur);
 256         switch (qp->s_state) {
 257         default:
 258                 /*
 259                  * Resend an old request or start a new one.
 260                  *
 261                  * We keep track of the current SWQE so that
 262                  * we don't reset the "furthest progress" state
 263                  * if we need to back up.
 264                  */
 265                 newreq = 0;
 266                 if (qp->s_cur == qp->s_tail) {
 267                         /* Check if send work queue is empty. */
 268                         if (qp->s_tail == qp->s_head)
 269                                 goto bail;
 270                         /*
 271                          * If a fence is requested, wait for previous
 272                          * RDMA read and atomic operations to finish.
 273                          */
 274                         if ((wqe->wr.send_flags & IB_SEND_FENCE) &&
 275                             qp->s_num_rd_atomic) {
 276                                 qp->s_flags |= IPATH_S_FENCE_PENDING;
 277                                 goto bail;
 278                         }
 279                         wqe->psn = qp->s_next_psn;
 280                         newreq = 1;
 281                 }
 282                 /*
 283                  * Note that we have to be careful not to modify the
 284                  * original work request since we may need to resend
 285                  * it.
 286                  */
 287                 len = wqe->length;
 288                 ss = &qp->s_sge;
 289                 bth2 = 0;
 290                 switch (wqe->wr.opcode) {
 291                 case IB_WR_SEND:
 292                 case IB_WR_SEND_WITH_IMM:
 293                         /* If no credit, return. */
 294                         if (qp->s_lsn != (u32) -1 &&
 295                             ipath_cmp24(wqe->ssn, qp->s_lsn + 1) > 0)
 296                                 goto bail;
 297                         wqe->lpsn = wqe->psn;
 298                         if (len > pmtu) {
 299                                 wqe->lpsn += (len - 1) / pmtu;
 300                                 qp->s_state = OP(SEND_FIRST);
 301                                 len = pmtu;
 302                                 break;
 303                         }
 304                         if (wqe->wr.opcode == IB_WR_SEND)
 305                                 qp->s_state = OP(SEND_ONLY);
 306                         else {
 307                                 qp->s_state = OP(SEND_ONLY_WITH_IMMEDIATE);
 308                                 /* Immediate data comes after the BTH */
 309                                 ohdr->u.imm_data = wqe->wr.imm_data;
 310                                 hwords += 1;
 311                         }
 312                         if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 313                                 bth0 |= 1 << 23;
 314                         bth2 = 1 << 31; /* Request ACK. */
 315                         if (++qp->s_cur == qp->s_size)
 316                                 qp->s_cur = 0;
 317                         break;
 318
 319                 case IB_WR_RDMA_WRITE:
 320                         if (newreq && qp->s_lsn != (u32) -1)
 321                                 qp->s_lsn++;
 322                         /* FALLTHROUGH */
 323                 case IB_WR_RDMA_WRITE_WITH_IMM:
 324                         /* If no credit, return. */
 325                         if (qp->s_lsn != (u32) -1 &&
 326                             ipath_cmp24(wqe->ssn, qp->s_lsn + 1) > 0)
 327                                 goto bail;
 328                         ohdr->u.rc.reth.vaddr =
 329                                 cpu_to_be64(wqe->wr.wr.rdma.remote_addr);
 330                         ohdr->u.rc.reth.rkey =
 331                                 cpu_to_be32(wqe->wr.wr.rdma.rkey);
 332                         ohdr->u.rc.reth.length = cpu_to_be32(len);
 333                         hwords += sizeof(struct ib_reth) / sizeof(u32);
 334                         wqe->lpsn = wqe->psn;
 335                         if (len > pmtu) {
 336                                 wqe->lpsn += (len - 1) / pmtu;
 337                                 qp->s_state = OP(RDMA_WRITE_FIRST);
 338                                 len = pmtu;
 339                                 break;
 340                         }
 341                         if (wqe->wr.opcode == IB_WR_RDMA_WRITE)
 342                                 qp->s_state = OP(RDMA_WRITE_ONLY);
 343                         else {
 344                                 qp->s_state =
 345                                         OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE);
 346                                 /* Immediate data comes after RETH */
 347                                 ohdr->u.rc.imm_data = wqe->wr.imm_data;
 348                                 hwords += 1;
 349                                 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 350                                         bth0 |= 1 << 23;
 351                         }
 352                         bth2 = 1 << 31; /* Request ACK. */
 353                         if (++qp->s_cur == qp->s_size)
 354                                 qp->s_cur = 0;
 355                         break;
 356
 357                 case IB_WR_RDMA_READ:
 358                         /*
 359                          * Don't allow more operations to be started
 360                          * than the QP limits allow.
 361                          */
 362                         if (newreq) {
 363                                 if (qp->s_num_rd_atomic >=
 364                                     qp->s_max_rd_atomic) {
 365                                         qp->s_flags |= IPATH_S_RDMAR_PENDING;
 366                                         goto bail;
 367                                 }
 368                                 qp->s_num_rd_atomic++;
 369                                 if (qp->s_lsn != (u32) -1)
 370                                         qp->s_lsn++;
 371                                 /*
 372                                  * Adjust s_next_psn to count the
 373                                  * expected number of responses.
 374                                  */
 375                                 if (len > pmtu)
 376                                         qp->s_next_psn += (len - 1) / pmtu;
 377                                 wqe->lpsn = qp->s_next_psn++;
 378                         }
 379                         ohdr->u.rc.reth.vaddr =
 380                                 cpu_to_be64(wqe->wr.wr.rdma.remote_addr);
 381                         ohdr->u.rc.reth.rkey =
 382                                 cpu_to_be32(wqe->wr.wr.rdma.rkey);
 383                         ohdr->u.rc.reth.length = cpu_to_be32(len);
 384                         qp->s_state = OP(RDMA_READ_REQUEST);
 385                         hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
 386                         ss = NULL;
 387                         len = 0;
 388                         if (++qp->s_cur == qp->s_size)
 389                                 qp->s_cur = 0;
 390                         break;
 391
 392                 case IB_WR_ATOMIC_CMP_AND_SWP:
 393                 case IB_WR_ATOMIC_FETCH_AND_ADD:
 394                         /*
 395                          * Don't allow more operations to be started
 396                          * than the QP limits allow.
 397                          */
 398                         if (newreq) {
 399                                 if (qp->s_num_rd_atomic >=
 400                                     qp->s_max_rd_atomic) {
 401                                         qp->s_flags |= IPATH_S_RDMAR_PENDING;
 402                                         goto bail;
 403                                 }
 404                                 qp->s_num_rd_atomic++;
 405                                 if (qp->s_lsn != (u32) -1)
 406                                         qp->s_lsn++;
 407                                 wqe->lpsn = wqe->psn;
 408                         }
 409                         if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
 410                                 qp->s_state = OP(COMPARE_SWAP);
 411                                 ohdr->u.atomic_eth.swap_data = cpu_to_be64(
 412                                         wqe->wr.wr.atomic.swap);
 413                                 ohdr->u.atomic_eth.compare_data = cpu_to_be64(
 414                                         wqe->wr.wr.atomic.compare_add);
 415                         } else {
 416                                 qp->s_state = OP(FETCH_ADD);
 417                                 ohdr->u.atomic_eth.swap_data = cpu_to_be64(
 418                                         wqe->wr.wr.atomic.compare_add);
 419                                 ohdr->u.atomic_eth.compare_data = 0;
 420                         }
 421                         ohdr->u.atomic_eth.vaddr[0] = cpu_to_be32(
 422                                 wqe->wr.wr.atomic.remote_addr >> 32);
 423                         ohdr->u.atomic_eth.vaddr[1] = cpu_to_be32(
 424                                 wqe->wr.wr.atomic.remote_addr);
 425                         ohdr->u.atomic_eth.rkey = cpu_to_be32(
 426                                 wqe->wr.wr.atomic.rkey);
 427                         hwords += sizeof(struct ib_atomic_eth) / sizeof(u32);
 428                         ss = NULL;
 429                         len = 0;
 430                         if (++qp->s_cur == qp->s_size)
 431                                 qp->s_cur = 0;
 432                         break;
 433
 434                 default:
 435                         goto bail;
 436                 }
 437                 qp->s_sge.sge = wqe->sg_list[0];
 438                 qp->s_sge.sg_list = wqe->sg_list + 1;
 439                 qp->s_sge.num_sge = wqe->wr.num_sge;
 440                 qp->s_len = wqe->length;
 441                 if (newreq) {
 442                         qp->s_tail++;
 443                         if (qp->s_tail >= qp->s_size)
 444                                 qp->s_tail = 0;
 445                 }
 446                 bth2 |= qp->s_psn & IPATH_PSN_MASK;
 447                 if (wqe->wr.opcode == IB_WR_RDMA_READ)
 448                         qp->s_psn = wqe->lpsn + 1;
 449                 else {
 450                         qp->s_psn++;
 451                         if (ipath_cmp24(qp->s_psn, qp->s_next_psn) > 0)
 452                                 qp->s_next_psn = qp->s_psn;
 453                 }
 454                 /*
 455                  * Put the QP on the pending list so lost ACKs will cause
 456                  * a retry.  More than one request can be pending so the
 457                  * QP may already be on the dev->pending list.
 458                  */
 459                 spin_lock(&dev->pending_lock);
 460                 if (list_empty(&qp->timerwait))
 461                         list_add_tail(&qp->timerwait,
 462                                       &dev->pending[dev->pending_index]);
 463                 spin_unlock(&dev->pending_lock);
 464                 break;
 465
 466         case OP(RDMA_READ_RESPONSE_FIRST):
 467                 /*
 468                  * This case can only happen if a send is restarted.
 469                  * See ipath_restart_rc().
 470                  */
 471                 ipath_init_restart(qp, wqe);
 472                 /* FALLTHROUGH */
 473         case OP(SEND_FIRST):
 474                 qp->s_state = OP(SEND_MIDDLE);
 475                 /* FALLTHROUGH */
 476         case OP(SEND_MIDDLE):
 477                 bth2 = qp->s_psn++ & IPATH_PSN_MASK;
 478                 if (ipath_cmp24(qp->s_psn, qp->s_next_psn) > 0)
 479                         qp->s_next_psn = qp->s_psn;
 480                 ss = &qp->s_sge;
 481                 len = qp->s_len;
 482                 if (len > pmtu) {
 483                         len = pmtu;
 484                         break;
 485                 }
 486                 if (wqe->wr.opcode == IB_WR_SEND)
 487                         qp->s_state = OP(SEND_LAST);
 488                 else {
 489                         qp->s_state = OP(SEND_LAST_WITH_IMMEDIATE);
 490                         /* Immediate data comes after the BTH */
 491                         ohdr->u.imm_data = wqe->wr.imm_data;
 492                         hwords += 1;
 493                 }
 494                 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 495                         bth0 |= 1 << 23;
 496                 bth2 |= 1 << 31;        /* Request ACK. */
 497                 qp->s_cur++;
 498                 if (qp->s_cur >= qp->s_size)
 499                         qp->s_cur = 0;
 500                 break;
 501
 502         case OP(RDMA_READ_RESPONSE_LAST):
 503                 /*
 504                  * This case can only happen if a RDMA write is restarted.
 505                  * See ipath_restart_rc().
 506                  */
 507                 ipath_init_restart(qp, wqe);
 508                 /* FALLTHROUGH */
 509         case OP(RDMA_WRITE_FIRST):
 510                 qp->s_state = OP(RDMA_WRITE_MIDDLE);
 511                 /* FALLTHROUGH */
 512         case OP(RDMA_WRITE_MIDDLE):
 513                 bth2 = qp->s_psn++ & IPATH_PSN_MASK;
 514                 if (ipath_cmp24(qp->s_psn, qp->s_next_psn) > 0)
 515                         qp->s_next_psn = qp->s_psn;
 516                 ss = &qp->s_sge;
 517                 len = qp->s_len;
 518                 if (len > pmtu) {
 519                         len = pmtu;
 520                         break;
 521                 }
 522                 if (wqe->wr.opcode == IB_WR_RDMA_WRITE)
 523                         qp->s_state = OP(RDMA_WRITE_LAST);
 524                 else {
 525                         qp->s_state = OP(RDMA_WRITE_LAST_WITH_IMMEDIATE);
 526                         /* Immediate data comes after the BTH */
 527                         ohdr->u.imm_data = wqe->wr.imm_data;
 528                         hwords += 1;
 529                         if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 530                                 bth0 |= 1 << 23;
 531                 }
 532                 bth2 |= 1 << 31;        /* Request ACK. */
 533                 qp->s_cur++;
 534                 if (qp->s_cur >= qp->s_size)
 535                         qp->s_cur = 0;
 536                 break;
 537
 538         case OP(RDMA_READ_RESPONSE_MIDDLE):
 539                 /*
 540                  * This case can only happen if a RDMA read is restarted.
 541                  * See ipath_restart_rc().
 542                  */
 543                 ipath_init_restart(qp, wqe);
 544                 len = ((qp->s_psn - wqe->psn) & IPATH_PSN_MASK) * pmtu;
 545                 ohdr->u.rc.reth.vaddr =
 546                         cpu_to_be64(wqe->wr.wr.rdma.remote_addr + len);
 547                 ohdr->u.rc.reth.rkey =
 548                         cpu_to_be32(wqe->wr.wr.rdma.rkey);
 549                 ohdr->u.rc.reth.length = cpu_to_be32(qp->s_len);
 550                 qp->s_state = OP(RDMA_READ_REQUEST);
 551                 hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
 552                 bth2 = qp->s_psn++ & IPATH_PSN_MASK;
 553                 if (ipath_cmp24(qp->s_psn, qp->s_next_psn) > 0)
 554                         qp->s_next_psn = qp->s_psn;
 555                 ss = NULL;
 556                 len = 0;
 557                 qp->s_cur++;
 558                 if (qp->s_cur == qp->s_size)
 559                         qp->s_cur = 0;
 560                 break;
 561         }
 562         if (ipath_cmp24(qp->s_psn, qp->s_last_psn + IPATH_PSN_CREDIT - 1) >= 0)
 563                 bth2 |= 1 << 31;        /* Request ACK. */
 564         qp->s_len -= len;
 565         qp->s_hdrwords = hwords;
 566         qp->s_cur_sge = ss;
 567         qp->s_cur_size = len;
 568         ipath_make_ruc_header(dev, qp, ohdr, bth0 | (qp->s_state << 24), bth2);
 569 done:
 570         ret = 1;
 571 bail:
 572         spin_unlock_irqrestore(&qp->s_lock, flags);
 573         return ret;
 574 }
 575
 576 /**
 577  * send_rc_ack - Construct an ACK packet and send it
 578  * @qp: a pointer to the QP
 579  *
 580  * This is called from ipath_rc_rcv() and only uses the receive
 581  * side QP state.
 582  * Note that RDMA reads and atomics are handled in the
 583  * send side QP state and tasklet.
 584  */
 585 static void send_rc_ack(struct ipath_qp *qp)
 586 {
 587         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
 588         u16 lrh0;
 589         u32 bth0;
 590         u32 hwords;
 591         struct ipath_ib_header hdr;
 592         struct ipath_other_headers *ohdr;
 593         unsigned long flags;
 594
 595         /* Don't send ACK or NAK if a RDMA read or atomic is pending. */
 596         if (qp->r_head_ack_queue != qp->s_tail_ack_queue ||
 597             (qp->s_flags & IPATH_S_ACK_PENDING) ||
 598             qp->s_ack_state != OP(ACKNOWLEDGE))
 599                 goto queue_ack;
 600
 601         /* Construct the header. */
 602         ohdr = &hdr.u.oth;
 603         lrh0 = IPATH_LRH_BTH;
 604         /* header size in 32-bit words LRH+BTH+AETH = (8+12+4)/4. */
 605         hwords = 6;
 606         if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
 607                 hwords += ipath_make_grh(dev, &hdr.u.l.grh,
 608                                          &qp->remote_ah_attr.grh,
 609                                          hwords, 0);
 610                 ohdr = &hdr.u.l.oth;
 611                 lrh0 = IPATH_LRH_GRH;
 612         }
 613         /* read pkey_index w/o lock (its atomic) */
 614         bth0 = ipath_get_pkey(dev->dd, qp->s_pkey_index) |
 615                 (OP(ACKNOWLEDGE) << 24) | (1 << 22);
 616         if (qp->r_nak_state)
 617                 ohdr->u.aeth = cpu_to_be32((qp->r_msn & IPATH_MSN_MASK) |
 618                                             (qp->r_nak_state <<
 619                                              IPATH_AETH_CREDIT_SHIFT));
 620         else
 621                 ohdr->u.aeth = ipath_compute_aeth(qp);
 622         lrh0 |= qp->remote_ah_attr.sl << 4;
 623         hdr.lrh[0] = cpu_to_be16(lrh0);
 624         hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
 625         hdr.lrh[2] = cpu_to_be16(hwords + SIZE_OF_CRC);
 626         hdr.lrh[3] = cpu_to_be16(dev->dd->ipath_lid);
 627         ohdr->bth[0] = cpu_to_be32(bth0);
 628         ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
 629         ohdr->bth[2] = cpu_to_be32(qp->r_ack_psn & IPATH_PSN_MASK);
 630
 631         /*
 632          * If we can send the ACK, clear the ACK state.
 633          */
 634         if (ipath_verbs_send(qp, &hdr, hwords, NULL, 0) == 0) {
 635                 dev->n_unicast_xmit++;
 636                 goto done;
 637         }
 638
 639         /*
 640          * We are out of PIO buffers at the moment.
 641          * Pass responsibility for sending the ACK to the
 642          * send tasklet so that when a PIO buffer becomes
 643          * available, the ACK is sent ahead of other outgoing
 644          * packets.
 645          */
 646         dev->n_rc_qacks++;
 647
 648 queue_ack:
 649         spin_lock_irqsave(&qp->s_lock, flags);
 650         qp->s_flags |= IPATH_S_ACK_PENDING;
 651         qp->s_nak_state = qp->r_nak_state;
 652         qp->s_ack_psn = qp->r_ack_psn;
 653         spin_unlock_irqrestore(&qp->s_lock, flags);
 654
 655         /* Call ipath_do_rc_send() in another thread. */
 656         tasklet_hi_schedule(&qp->s_task);
 657
 658 done:
 659         return;
 660 }
 661
 662 /**
 663  * reset_psn - reset the QP state to send starting from PSN
 664  * @qp: the QP
 665  * @psn: the packet sequence number to restart at
 666  *
 667  * This is called from ipath_rc_rcv() to process an incoming RC ACK
 668  * for the given QP.
 669  * Called at interrupt level with the QP s_lock held.
 670  */
 671 static void reset_psn(struct ipath_qp *qp, u32 psn)
 672 {
 673         u32 n = qp->s_last;
 674         struct ipath_swqe *wqe = get_swqe_ptr(qp, n);
 675         u32 opcode;
 676
 677         qp->s_cur = n;
 678
 679         /*
 680          * If we are starting the request from the beginning,
 681          * let the normal send code handle initialization.
 682          */
 683         if (ipath_cmp24(psn, wqe->psn) <= 0) {
 684                 qp->s_state = OP(SEND_LAST);
 685                 goto done;
 686         }
 687
 688         /* Find the work request opcode corresponding to the given PSN. */
 689         opcode = wqe->wr.opcode;
 690         for (;;) {
 691                 int diff;
 692
 693                 if (++n == qp->s_size)
 694                         n = 0;
 695                 if (n == qp->s_tail)
 696                         break;
 697                 wqe = get_swqe_ptr(qp, n);
 698                 diff = ipath_cmp24(psn, wqe->psn);
 699                 if (diff < 0)
 700                         break;
 701                 qp->s_cur = n;
 702                 /*
 703                  * If we are starting the request from the beginning,
 704                  * let the normal send code handle initialization.
 705                  */
 706                 if (diff == 0) {
 707                         qp->s_state = OP(SEND_LAST);
 708                         goto done;
 709                 }
 710                 opcode = wqe->wr.opcode;
 711         }
 712
 713         /*
 714          * Set the state to restart in the middle of a request.
 715          * Don't change the s_sge, s_cur_sge, or s_cur_size.
 716          * See ipath_do_rc_send().
 717          */
 718         switch (opcode) {
 719         case IB_WR_SEND:
 720         case IB_WR_SEND_WITH_IMM:
 721                 qp->s_state = OP(RDMA_READ_RESPONSE_FIRST);
 722                 break;
 723
 724         case IB_WR_RDMA_WRITE:
 725         case IB_WR_RDMA_WRITE_WITH_IMM:
 726                 qp->s_state = OP(RDMA_READ_RESPONSE_LAST);
 727                 break;
 728
 729         case IB_WR_RDMA_READ:
 730                 qp->s_state = OP(RDMA_READ_RESPONSE_MIDDLE);
 731                 break;
 732
 733         default:
 734                 /*
 735                  * This case shouldn't happen since its only
 736                  * one PSN per req.
 737                  */
 738                 qp->s_state = OP(SEND_LAST);
 739         }
 740 done:
 741         qp->s_psn = psn;
 742 }
 743
 744 /**
 745  * ipath_restart_rc - back up requester to resend the last un-ACKed request
 746  * @qp: the QP to restart
 747  * @psn: packet sequence number for the request
 748  * @wc: the work completion request
 749  *
 750  * The QP s_lock should be held and interrupts disabled.
 751  */
 752 void ipath_restart_rc(struct ipath_qp *qp, u32 psn, struct ib_wc *wc)
 753 {
 754         struct ipath_swqe *wqe = get_swqe_ptr(qp, qp->s_last);
 755         struct ipath_ibdev *dev;
 756
 757         if (qp->s_retry == 0) {
 758                 wc->wr_id = wqe->wr.wr_id;
 759                 wc->status = IB_WC_RETRY_EXC_ERR;
 760                 wc->opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
 761                 wc->vendor_err = 0;
 762                 wc->byte_len = 0;
 763                 wc->qp = &qp->ibqp;
 764                 wc->imm_data = 0;
 765                 wc->src_qp = qp->remote_qpn;
 766                 wc->wc_flags = 0;
 767                 wc->pkey_index = 0;
 768                 wc->slid = qp->remote_ah_attr.dlid;
 769                 wc->sl = qp->remote_ah_attr.sl;
 770                 wc->dlid_path_bits = 0;
 771                 wc->port_num = 0;
 772                 ipath_sqerror_qp(qp, wc);
 773                 goto bail;
 774         }
 775         qp->s_retry--;
 776
 777         /*
 778          * Remove the QP from the timeout queue.
 779          * Note: it may already have been removed by ipath_ib_timer().
 780          */
 781         dev = to_idev(qp->ibqp.device);
 782         spin_lock(&dev->pending_lock);
 783         if (!list_empty(&qp->timerwait))
 784                 list_del_init(&qp->timerwait);
 785         spin_unlock(&dev->pending_lock);
 786
 787         if (wqe->wr.opcode == IB_WR_RDMA_READ)
 788                 dev->n_rc_resends++;
 789         else
 790                 dev->n_rc_resends += (qp->s_psn - psn) & IPATH_PSN_MASK;
 791
 792         reset_psn(qp, psn);
 793         tasklet_hi_schedule(&qp->s_task);
 794
 795 bail:
 796         return;
 797 }
 798
 799 static inline void update_last_psn(struct ipath_qp *qp, u32 psn)
 800 {
 801         if (qp->s_wait_credit) {
 802                 qp->s_wait_credit = 0;
 803                 tasklet_hi_schedule(&qp->s_task);
 804         }
 805         qp->s_last_psn = psn;
 806 }
 807
 808 /**
 809  * do_rc_ack - process an incoming RC ACK
 810  * @qp: the QP the ACK came in on
 811  * @psn: the packet sequence number of the ACK
 812  * @opcode: the opcode of the request that resulted in the ACK
 813  *
 814  * This is called from ipath_rc_rcv_resp() to process an incoming RC ACK
 815  * for the given QP.
 816  * Called at interrupt level with the QP s_lock held and interrupts disabled.
 817  * Returns 1 if OK, 0 if current operation should be aborted (NAK).
 818  */
 819 static int do_rc_ack(struct ipath_qp *qp, u32 aeth, u32 psn, int opcode,
 820                      u64 val)
 821 {
 822         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
 823         struct ib_wc wc;
 824         struct ipath_swqe *wqe;
 825         int ret = 0;
 826         u32 ack_psn;
 827         int diff;
 828
 829         /*
 830          * Remove the QP from the timeout queue (or RNR timeout queue).
 831          * If ipath_ib_timer() has already removed it,
 832          * it's OK since we hold the QP s_lock and ipath_restart_rc()
 833          * just won't find anything to restart if we ACK everything.
 834          */
 835         spin_lock(&dev->pending_lock);
 836         if (!list_empty(&qp->timerwait))
 837                 list_del_init(&qp->timerwait);
 838         spin_unlock(&dev->pending_lock);
 839
 840         /*
 841          * Note that NAKs implicitly ACK outstanding SEND and RDMA write
 842          * requests and implicitly NAK RDMA read and atomic requests issued
 843          * before the NAK'ed request.  The MSN won't include the NAK'ed
 844          * request but will include an ACK'ed request(s).
 845          */
 846         ack_psn = psn;
 847         if (aeth >> 29)
 848                 ack_psn--;
 849         wqe = get_swqe_ptr(qp, qp->s_last);
 850
 851         /*
 852          * The MSN might be for a later WQE than the PSN indicates so
 853          * only complete WQEs that the PSN finishes.
 854          */
 855         while ((diff = ipath_cmp24(ack_psn, wqe->lpsn)) >= 0) {
 856                 /*
 857                  * RDMA_READ_RESPONSE_ONLY is a special case since
 858                  * we want to generate completion events for everything
 859                  * before the RDMA read, copy the data, then generate
 860                  * the completion for the read.
 861                  */
 862                 if (wqe->wr.opcode == IB_WR_RDMA_READ &&
 863                     opcode == OP(RDMA_READ_RESPONSE_ONLY) &&
 864                     diff == 0) {
 865                         ret = 1;
 866                         goto bail;
 867                 }
 868                 /*
 869                  * If this request is a RDMA read or atomic, and the ACK is
 870                  * for a later operation, this ACK NAKs the RDMA read or
 871                  * atomic.  In other words, only a RDMA_READ_LAST or ONLY
 872                  * can ACK a RDMA read and likewise for atomic ops.  Note
 873                  * that the NAK case can only happen if relaxed ordering is
 874                  * used and requests are sent after an RDMA read or atomic
 875                  * is sent but before the response is received.
 876                  */
 877                 if ((wqe->wr.opcode == IB_WR_RDMA_READ &&
 878                      (opcode != OP(RDMA_READ_RESPONSE_LAST) || diff != 0)) ||
 879                     ((wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
 880                       wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) &&
 881                      (opcode != OP(ATOMIC_ACKNOWLEDGE) || diff != 0))) {
 882                         /*
 883                          * The last valid PSN seen is the previous
 884                          * request's.
 885                          */
 886                         update_last_psn(qp, wqe->psn - 1);
 887                         /* Retry this request. */
 888                         ipath_restart_rc(qp, wqe->psn, &wc);
 889                         /*
 890                          * No need to process the ACK/NAK since we are
 891                          * restarting an earlier request.
 892                          */
 893                         goto bail;
 894                 }
 895                 if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
 896                     wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)
 897                         *(u64 *) wqe->sg_list[0].vaddr = val;
 898                 if (qp->s_num_rd_atomic &&
 899                     (wqe->wr.opcode == IB_WR_RDMA_READ ||
 900                      wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
 901                      wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)) {
 902                         qp->s_num_rd_atomic--;
 903                         /* Restart sending task if fence is complete */
 904                         if ((qp->s_flags & IPATH_S_FENCE_PENDING) &&
 905                             !qp->s_num_rd_atomic) {
 906                                 qp->s_flags &= ~IPATH_S_FENCE_PENDING;
 907                                 tasklet_hi_schedule(&qp->s_task);
 908                         } else if (qp->s_flags & IPATH_S_RDMAR_PENDING) {
 909                                 qp->s_flags &= ~IPATH_S_RDMAR_PENDING;
 910                                 tasklet_hi_schedule(&qp->s_task);
 911                         }
 912                 }
 913                 /* Post a send completion queue entry if requested. */
 914                 if (!(qp->s_flags & IPATH_S_SIGNAL_REQ_WR) ||
 915                     (wqe->wr.send_flags & IB_SEND_SIGNALED)) {
 916                         wc.wr_id = wqe->wr.wr_id;
 917                         wc.status = IB_WC_SUCCESS;
 918                         wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
 919                         wc.vendor_err = 0;
 920                         wc.byte_len = wqe->length;
 921                         wc.imm_data = 0;
 922                         wc.qp = &qp->ibqp;
 923                         wc.src_qp = qp->remote_qpn;
 924                         wc.wc_flags = 0;
 925                         wc.pkey_index = 0;
 926                         wc.slid = qp->remote_ah_attr.dlid;
 927                         wc.sl = qp->remote_ah_attr.sl;
 928                         wc.dlid_path_bits = 0;
 929                         wc.port_num = 0;
 930                         ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 0);
 931                 }
 932                 qp->s_retry = qp->s_retry_cnt;
 933                 /*
 934                  * If we are completing a request which is in the process of
 935                  * being resent, we can stop resending it since we know the
 936                  * responder has already seen it.
 937                  */
 938                 if (qp->s_last == qp->s_cur) {
 939                         if (++qp->s_cur >= qp->s_size)
 940                                 qp->s_cur = 0;
 941                         qp->s_last = qp->s_cur;
 942                         if (qp->s_last == qp->s_tail)
 943                                 break;
 944                         wqe = get_swqe_ptr(qp, qp->s_cur);
 945                         qp->s_state = OP(SEND_LAST);
 946                         qp->s_psn = wqe->psn;
 947                 } else {
 948                         if (++qp->s_last >= qp->s_size)
 949                                 qp->s_last = 0;
 950                         if (qp->s_last == qp->s_tail)
 951                                 break;
 952                         wqe = get_swqe_ptr(qp, qp->s_last);
 953                 }
 954         }
 955
 956         switch (aeth >> 29) {
 957         case 0:         /* ACK */
 958                 dev->n_rc_acks++;
 959                 /* If this is a partial ACK, reset the retransmit timer. */
 960                 if (qp->s_last != qp->s_tail) {
 961                         spin_lock(&dev->pending_lock);
 962                         if (list_empty(&qp->timerwait))
 963                                 list_add_tail(&qp->timerwait,
 964                                         &dev->pending[dev->pending_index]);
 965                         spin_unlock(&dev->pending_lock);
 966                         /*
 967                          * If we get a partial ACK for a resent operation,
 968                          * we can stop resending the earlier packets and
 969                          * continue with the next packet the receiver wants.
 970                          */
 971                         if (ipath_cmp24(qp->s_psn, psn) <= 0) {
 972                                 reset_psn(qp, psn + 1);
 973                                 tasklet_hi_schedule(&qp->s_task);
 974                         }
 975                 } else if (ipath_cmp24(qp->s_psn, psn) <= 0) {
 976                         qp->s_state = OP(SEND_LAST);
 977                         qp->s_psn = psn + 1;
 978                 }
 979                 ipath_get_credit(qp, aeth);
 980                 qp->s_rnr_retry = qp->s_rnr_retry_cnt;
 981                 qp->s_retry = qp->s_retry_cnt;
 982                 update_last_psn(qp, psn);
 983                 ret = 1;
 984                 goto bail;
 985
 986         case 1:         /* RNR NAK */
 987                 dev->n_rnr_naks++;
 988                 if (qp->s_last == qp->s_tail)
 989                         goto bail;
 990                 if (qp->s_rnr_retry == 0) {
 991                         wc.status = IB_WC_RNR_RETRY_EXC_ERR;
 992                         goto class_b;
 993                 }
 994                 if (qp->s_rnr_retry_cnt < 7)
 995                         qp->s_rnr_retry--;
 996
 997                 /* The last valid PSN is the previous PSN. */
 998                 update_last_psn(qp, psn - 1);
 999
1000                 if (wqe->wr.opcode == IB_WR_RDMA_READ)
1001                         dev->n_rc_resends++;
1002                 else
1003                         dev->n_rc_resends +=
1004                                 (qp->s_psn - psn) & IPATH_PSN_MASK;
1005
1006                 reset_psn(qp, psn);
1007
1008                 qp->s_rnr_timeout =
1009                         ib_ipath_rnr_table[(aeth >> IPATH_AETH_CREDIT_SHIFT) &
1010                                            IPATH_AETH_CREDIT_MASK];
1011                 ipath_insert_rnr_queue(qp);
1012                 goto bail;
1013
1014         case 3:         /* NAK */
1015                 if (qp->s_last == qp->s_tail)
1016                         goto bail;
1017                 /* The last valid PSN is the previous PSN. */
1018                 update_last_psn(qp, psn - 1);
1019                 switch ((aeth >> IPATH_AETH_CREDIT_SHIFT) &
1020                         IPATH_AETH_CREDIT_MASK) {
1021                 case 0: /* PSN sequence error */
1022                         dev->n_seq_naks++;
1023                         /*
1024                          * Back up to the responder's expected PSN.
1025                          * Note that we might get a NAK in the middle of an
1026                          * RDMA READ response which terminates the RDMA
1027                          * READ.
1028                          */
1029                         ipath_restart_rc(qp, psn, &wc);
1030                         break;
1031
1032                 case 1: /* Invalid Request */
1033                         wc.status = IB_WC_REM_INV_REQ_ERR;
1034                         dev->n_other_naks++;
1035                         goto class_b;
1036
1037                 case 2: /* Remote Access Error */
1038                         wc.status = IB_WC_REM_ACCESS_ERR;
1039                         dev->n_other_naks++;
1040                         goto class_b;
1041
1042                 case 3: /* Remote Operation Error */
1043                         wc.status = IB_WC_REM_OP_ERR;
1044                         dev->n_other_naks++;
1045                 class_b:
1046                         wc.wr_id = wqe->wr.wr_id;
1047                         wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
1048                         wc.vendor_err = 0;
1049                         wc.byte_len = 0;
1050                         wc.qp = &qp->ibqp;
1051                         wc.imm_data = 0;
1052                         wc.src_qp = qp->remote_qpn;
1053                         wc.wc_flags = 0;
1054                         wc.pkey_index = 0;
1055                         wc.slid = qp->remote_ah_attr.dlid;
1056                         wc.sl = qp->remote_ah_attr.sl;
1057                         wc.dlid_path_bits = 0;
1058                         wc.port_num = 0;
1059                         ipath_sqerror_qp(qp, &wc);
1060                         break;
1061
1062                 default:
1063                         /* Ignore other reserved NAK error codes */
1064                         goto reserved;
1065                 }
1066                 qp->s_rnr_retry = qp->s_rnr_retry_cnt;
1067                 goto bail;
1068
1069         default:                /* 2: reserved */
1070         reserved:
1071                 /* Ignore reserved NAK codes. */
1072                 goto bail;
1073         }
1074
1075 bail:
1076         return ret;
1077 }
1078
1079 /**
1080  * ipath_rc_rcv_resp - process an incoming RC response packet
1081  * @dev: the device this packet came in on
1082  * @ohdr: the other headers for this packet
1083  * @data: the packet data
1084  * @tlen: the packet length
1085  * @qp: the QP for this packet
1086  * @opcode: the opcode for this packet
1087  * @psn: the packet sequence number for this packet
1088  * @hdrsize: the header length
1089  * @pmtu: the path MTU
1090  * @header_in_data: true if part of the header data is in the data buffer
1091  *
1092  * This is called from ipath_rc_rcv() to process an incoming RC response
1093  * packet for the given QP.
1094  * Called at interrupt level.
1095  */
1096 static inline void ipath_rc_rcv_resp(struct ipath_ibdev *dev,
1097                                      struct ipath_other_headers *ohdr,
1098                                      void *data, u32 tlen,
1099                                      struct ipath_qp *qp,
1100                                      u32 opcode,
1101                                      u32 psn, u32 hdrsize, u32 pmtu,
1102                                      int header_in_data)
1103 {
1104         struct ipath_swqe *wqe;
1105         unsigned long flags;
1106         struct ib_wc wc;
1107         int diff;
1108         u32 pad;
1109         u32 aeth;
1110         u64 val;
1111
1112         spin_lock_irqsave(&qp->s_lock, flags);
1113
1114         /* Ignore invalid responses. */
1115         if (ipath_cmp24(psn, qp->s_next_psn) >= 0)
1116                 goto ack_done;
1117
1118         /* Ignore duplicate responses. */
1119         diff = ipath_cmp24(psn, qp->s_last_psn);
1120         if (unlikely(diff <= 0)) {
1121                 /* Update credits for "ghost" ACKs */
1122                 if (diff == 0 && opcode == OP(ACKNOWLEDGE)) {
1123                         if (!header_in_data)
1124                                 aeth = be32_to_cpu(ohdr->u.aeth);
1125                         else {
1126                                 aeth = be32_to_cpu(((__be32 *) data)[0]);
1127                                 data += sizeof(__be32);
1128                         }
1129                         if ((aeth >> 29) == 0)
1130                                 ipath_get_credit(qp, aeth);
1131                 }
1132                 goto ack_done;
1133         }
1134
1135         if (unlikely(qp->s_last == qp->s_tail))
1136                 goto ack_done;
1137         wqe = get_swqe_ptr(qp, qp->s_last);
1138
1139         switch (opcode) {
1140         case OP(ACKNOWLEDGE):
1141         case OP(ATOMIC_ACKNOWLEDGE):
1142         case OP(RDMA_READ_RESPONSE_FIRST):
1143                 if (!header_in_data)
1144                         aeth = be32_to_cpu(ohdr->u.aeth);
1145                 else {
1146                         aeth = be32_to_cpu(((__be32 *) data)[0]);
1147                         data += sizeof(__be32);
1148                 }
1149                 if (opcode == OP(ATOMIC_ACKNOWLEDGE)) {
1150                         if (!header_in_data) {
1151                                 __be32 *p = ohdr->u.at.atomic_ack_eth;
1152
1153                                 val = ((u64) be32_to_cpu(p[0]) << 32) |
1154                                         be32_to_cpu(p[1]);
1155                         } else
1156                                 val = be64_to_cpu(((__be64 *) data)[0]);
1157                 } else
1158                         val = 0;
1159                 if (!do_rc_ack(qp, aeth, psn, opcode, val) ||
1160                     opcode != OP(RDMA_READ_RESPONSE_FIRST))
1161                         goto ack_done;
1162                 hdrsize += 4;
1163                 wqe = get_swqe_ptr(qp, qp->s_last);
1164                 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1165                         goto ack_op_err;
1166                 /*
1167                  * If this is a response to a resent RDMA read, we
1168                  * have to be careful to copy the data to the right
1169                  * location.
1170                  */
1171                 qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
1172                                                   wqe, psn, pmtu);
1173                 goto read_middle;
1174
1175         case OP(RDMA_READ_RESPONSE_MIDDLE):
1176                 /* no AETH, no ACK */
1177                 if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) {
1178                         dev->n_rdma_seq++;
1179                         ipath_restart_rc(qp, qp->s_last_psn + 1, &wc);
1180                         goto ack_done;
1181                 }
1182                 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1183                         goto ack_op_err;
1184         read_middle:
1185                 if (unlikely(tlen != (hdrsize + pmtu + 4)))
1186                         goto ack_len_err;
1187                 if (unlikely(pmtu >= qp->s_rdma_read_len))
1188                         goto ack_len_err;
1189
1190                 /* We got a response so update the timeout. */
1191                 spin_lock(&dev->pending_lock);
1192                 if (qp->s_rnr_timeout == 0 && !list_empty(&qp->timerwait))
1193                         list_move_tail(&qp->timerwait,
1194                                        &dev->pending[dev->pending_index]);
1195                 spin_unlock(&dev->pending_lock);
1196                 /*
1197                  * Update the RDMA receive state but do the copy w/o
1198                  * holding the locks and blocking interrupts.
1199                  */
1200                 qp->s_rdma_read_len -= pmtu;
1201                 update_last_psn(qp, psn);
1202                 spin_unlock_irqrestore(&qp->s_lock, flags);
1203                 ipath_copy_sge(&qp->s_rdma_read_sge, data, pmtu);
1204                 goto bail;
1205
1206         case OP(RDMA_READ_RESPONSE_ONLY):
1207                 if (!header_in_data)
1208                         aeth = be32_to_cpu(ohdr->u.aeth);
1209                 else
1210                         aeth = be32_to_cpu(((__be32 *) data)[0]);
1211                 if (!do_rc_ack(qp, aeth, psn, opcode, 0))
1212                         goto ack_done;
1213                 /* Get the number of bytes the message was padded by. */
1214                 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1215                 /*
1216                  * Check that the data size is >= 0 && <= pmtu.
1217                  * Remember to account for the AETH header (4) and
1218                  * ICRC (4).
1219                  */
1220                 if (unlikely(tlen < (hdrsize + pad + 8)))
1221                         goto ack_len_err;
1222                 /*
1223                  * If this is a response to a resent RDMA read, we
1224                  * have to be careful to copy the data to the right
1225                  * location.
1226                  */
1227                 wqe = get_swqe_ptr(qp, qp->s_last);
1228                 qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
1229                                                   wqe, psn, pmtu);
1230                 goto read_last;
1231
1232         case OP(RDMA_READ_RESPONSE_LAST):
1233                 /* ACKs READ req. */
1234                 if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) {
1235                         dev->n_rdma_seq++;
1236                         ipath_restart_rc(qp, qp->s_last_psn + 1, &wc);
1237                         goto ack_done;
1238                 }
1239                 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1240                         goto ack_op_err;
1241                 /* Get the number of bytes the message was padded by. */
1242                 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1243                 /*
1244                  * Check that the data size is >= 1 && <= pmtu.
1245                  * Remember to account for the AETH header (4) and
1246                  * ICRC (4).
1247                  */
1248                 if (unlikely(tlen <= (hdrsize + pad + 8)))
1249                         goto ack_len_err;
1250         read_last:
1251                 tlen -= hdrsize + pad + 8;
1252                 if (unlikely(tlen != qp->s_rdma_read_len))
1253                         goto ack_len_err;
1254                 if (!header_in_data)
1255                         aeth = be32_to_cpu(ohdr->u.aeth);
1256                 else {
1257                         aeth = be32_to_cpu(((__be32 *) data)[0]);
1258                         data += sizeof(__be32);
1259                 }
1260                 ipath_copy_sge(&qp->s_rdma_read_sge, data, tlen);
1261                 (void) do_rc_ack(qp, aeth, psn,
1262                                  OP(RDMA_READ_RESPONSE_LAST), 0);
1263                 goto ack_done;
1264         }
1265
1266 ack_done:
1267         spin_unlock_irqrestore(&qp->s_lock, flags);
1268         goto bail;
1269
1270 ack_op_err:
1271         wc.status = IB_WC_LOC_QP_OP_ERR;
1272         goto ack_err;
1273
1274 ack_len_err:
1275         wc.status = IB_WC_LOC_LEN_ERR;
1276 ack_err:
1277         wc.wr_id = wqe->wr.wr_id;
1278         wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
1279         wc.vendor_err = 0;
1280         wc.byte_len = 0;
1281         wc.imm_data = 0;
1282         wc.qp = &qp->ibqp;
1283         wc.src_qp = qp->remote_qpn;
1284         wc.wc_flags = 0;
1285         wc.pkey_index = 0;
1286         wc.slid = qp->remote_ah_attr.dlid;
1287         wc.sl = qp->remote_ah_attr.sl;
1288         wc.dlid_path_bits = 0;
1289         wc.port_num = 0;
1290         ipath_sqerror_qp(qp, &wc);
1291         spin_unlock_irqrestore(&qp->s_lock, flags);
1292 bail:
1293         return;
1294 }
1295
1296 /**
1297  * ipath_rc_rcv_error - process an incoming duplicate or error RC packet
1298  * @dev: the device this packet came in on
1299  * @ohdr: the other headers for this packet
1300  * @data: the packet data
1301  * @qp: the QP for this packet
1302  * @opcode: the opcode for this packet
1303  * @psn: the packet sequence number for this packet
1304  * @diff: the difference between the PSN and the expected PSN
1305  * @header_in_data: true if part of the header data is in the data buffer
1306  *
1307  * This is called from ipath_rc_rcv() to process an unexpected
1308  * incoming RC packet for the given QP.
1309  * Called at interrupt level.
1310  * Return 1 if no more processing is needed; otherwise return 0 to
1311  * schedule a response to be sent.
1312  */
1313 static inline int ipath_rc_rcv_error(struct ipath_ibdev *dev,
1314                                      struct ipath_other_headers *ohdr,
1315                                      void *data,
1316                                      struct ipath_qp *qp,
1317                                      u32 opcode,
1318                                      u32 psn,
1319                                      int diff,
1320                                      int header_in_data)
1321 {
1322         struct ipath_ack_entry *e;
1323         u8 i, prev;
1324         int old_req;
1325         unsigned long flags;
1326
1327         if (diff > 0) {
1328                 /*
1329                  * Packet sequence error.
1330                  * A NAK will ACK earlier sends and RDMA writes.
1331                  * Don't queue the NAK if we already sent one.
1332                  */
1333                 if (!qp->r_nak_state) {
1334                         qp->r_nak_state = IB_NAK_PSN_ERROR;
1335                         /* Use the expected PSN. */
1336                         qp->r_ack_psn = qp->r_psn;
1337                         goto send_ack;
1338                 }
1339                 goto done;
1340         }
1341
1342         /*
1343          * Handle a duplicate request.  Don't re-execute SEND, RDMA
1344          * write or atomic op.  Don't NAK errors, just silently drop
1345          * the duplicate request.  Note that r_sge, r_len, and
1346          * r_rcv_len may be in use so don't modify them.
1347          *
1348          * We are supposed to ACK the earliest duplicate PSN but we
1349          * can coalesce an outstanding duplicate ACK.  We have to
1350          * send the earliest so that RDMA reads can be restarted at
1351          * the requester's expected PSN.
1352          *
1353          * First, find where this duplicate PSN falls within the
1354          * ACKs previously sent.
1355          */
1356         psn &= IPATH_PSN_MASK;
1357         e = NULL;
1358         old_req = 1;
1359         spin_lock_irqsave(&qp->s_lock, flags);
1360         for (i = qp->r_head_ack_queue; ; i = prev) {
1361                 if (i == qp->s_tail_ack_queue)
1362                         old_req = 0;
1363                 if (i)
1364                         prev = i - 1;
1365                 else
1366                         prev = IPATH_MAX_RDMA_ATOMIC;
1367                 if (prev == qp->r_head_ack_queue) {
1368                         e = NULL;
1369                         break;
1370                 }
1371                 e = &qp->s_ack_queue[prev];
1372                 if (!e->opcode) {
1373                         e = NULL;
1374                         break;
1375                 }
1376                 if (ipath_cmp24(psn, e->psn) >= 0) {
1377                         if (prev == qp->s_tail_ack_queue)
1378                                 old_req = 0;
1379                         break;
1380                 }
1381         }
1382         switch (opcode) {
1383         case OP(RDMA_READ_REQUEST): {
1384                 struct ib_reth *reth;
1385                 u32 offset;
1386                 u32 len;
1387
1388                 /*
1389                  * If we didn't find the RDMA read request in the ack queue,
1390                  * or the send tasklet is already backed up to send an
1391                  * earlier entry, we can ignore this request.
1392                  */
1393                 if (!e || e->opcode != OP(RDMA_READ_REQUEST) || old_req)
1394                         goto unlock_done;
1395                 /* RETH comes after BTH */
1396                 if (!header_in_data)
1397                         reth = &ohdr->u.rc.reth;
1398                 else {
1399                         reth = (struct ib_reth *)data;
1400                         data += sizeof(*reth);
1401                 }
1402                 /*
1403                  * Address range must be a subset of the original
1404                  * request and start on pmtu boundaries.
1405                  * We reuse the old ack_queue slot since the requester
1406                  * should not back up and request an earlier PSN for the
1407                  * same request.
1408                  */
1409                 offset = ((psn - e->psn) & IPATH_PSN_MASK) *
1410                         ib_mtu_enum_to_int(qp->path_mtu);
1411                 len = be32_to_cpu(reth->length);
1412                 if (unlikely(offset + len > e->rdma_sge.sge.sge_length))
1413                         goto unlock_done;
1414                 if (len != 0) {
1415                         u32 rkey = be32_to_cpu(reth->rkey);
1416                         u64 vaddr = be64_to_cpu(reth->vaddr);
1417                         int ok;
1418
1419                         ok = ipath_rkey_ok(qp, &e->rdma_sge,
1420                                            len, vaddr, rkey,
1421                                            IB_ACCESS_REMOTE_READ);
1422                         if (unlikely(!ok))
1423                                 goto unlock_done;
1424                 } else {
1425                         e->rdma_sge.sg_list = NULL;
1426                         e->rdma_sge.num_sge = 0;
1427                         e->rdma_sge.sge.mr = NULL;
1428                         e->rdma_sge.sge.vaddr = NULL;
1429                         e->rdma_sge.sge.length = 0;
1430                         e->rdma_sge.sge.sge_length = 0;
1431                 }
1432                 e->psn = psn;
1433                 qp->s_ack_state = OP(ACKNOWLEDGE);
1434                 qp->s_tail_ack_queue = prev;
1435                 break;
1436         }
1437
1438         case OP(COMPARE_SWAP):
1439         case OP(FETCH_ADD): {
1440                 /*
1441                  * If we didn't find the atomic request in the ack queue
1442                  * or the send tasklet is already backed up to send an
1443                  * earlier entry, we can ignore this request.
1444                  */
1445                 if (!e || e->opcode != (u8) opcode || old_req)
1446                         goto unlock_done;
1447                 qp->s_ack_state = OP(ACKNOWLEDGE);
1448                 qp->s_tail_ack_queue = prev;
1449                 break;
1450         }
1451
1452         default:
1453                 if (old_req)
1454                         goto unlock_done;
1455                 /*
1456                  * Resend the most recent ACK if this request is
1457                  * after all the previous RDMA reads and atomics.
1458                  */
1459                 if (i == qp->r_head_ack_queue) {
1460                         spin_unlock_irqrestore(&qp->s_lock, flags);
1461                         qp->r_nak_state = 0;
1462                         qp->r_ack_psn = qp->r_psn - 1;
1463                         goto send_ack;
1464                 }
1465                 /*
1466                  * Try to send a simple ACK to work around a Mellanox bug
1467                  * which doesn't accept a RDMA read response or atomic
1468                  * response as an ACK for earlier SENDs or RDMA writes.
1469                  */
1470                 if (qp->r_head_ack_queue == qp->s_tail_ack_queue &&
1471                     !(qp->s_flags & IPATH_S_ACK_PENDING) &&
1472                     qp->s_ack_state == OP(ACKNOWLEDGE)) {
1473                         spin_unlock_irqrestore(&qp->s_lock, flags);
1474                         qp->r_nak_state = 0;
1475                         qp->r_ack_psn = qp->s_ack_queue[i].psn - 1;
1476                         goto send_ack;
1477                 }
1478                 /*
1479                  * Resend the RDMA read or atomic op which
1480                  * ACKs this duplicate request.
1481                  */
1482                 qp->s_ack_state = OP(ACKNOWLEDGE);
1483                 qp->s_tail_ack_queue = i;
1484                 break;
1485         }
1486         qp->r_nak_state = 0;
1487         tasklet_hi_schedule(&qp->s_task);
1488
1489 unlock_done:
1490         spin_unlock_irqrestore(&qp->s_lock, flags);
1491 done:
1492         return 1;
1493
1494 send_ack:
1495         return 0;
1496 }
1497
1498 static void ipath_rc_error(struct ipath_qp *qp, enum ib_wc_status err)
1499 {
1500         unsigned long flags;
1501         int lastwqe;
1502
1503         spin_lock_irqsave(&qp->s_lock, flags);
1504         qp->state = IB_QPS_ERR;
1505         lastwqe = ipath_error_qp(qp, err);
1506         spin_unlock_irqrestore(&qp->s_lock, flags);
1507
1508         if (lastwqe) {
1509                 struct ib_event ev;
1510
1511                 ev.device = qp->ibqp.device;
1512                 ev.element.qp = &qp->ibqp;
1513                 ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
1514                 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
1515         }
1516 }
1517
1518 static inline void ipath_update_ack_queue(struct ipath_qp *qp, unsigned n)
1519 {
1520         unsigned long flags;
1521         unsigned next;
1522
1523         next = n + 1;
1524         if (next > IPATH_MAX_RDMA_ATOMIC)
1525                 next = 0;
1526         spin_lock_irqsave(&qp->s_lock, flags);
1527         if (n == qp->s_tail_ack_queue) {
1528                 qp->s_tail_ack_queue = next;
1529                 qp->s_ack_state = OP(ACKNOWLEDGE);
1530         }
1531         spin_unlock_irqrestore(&qp->s_lock, flags);
1532 }
1533
1534 /**
1535  * ipath_rc_rcv - process an incoming RC packet
1536  * @dev: the device this packet came in on
1537  * @hdr: the header of this packet
1538  * @has_grh: true if the header has a GRH
1539  * @data: the packet data
1540  * @tlen: the packet length
1541  * @qp: the QP for this packet
1542  *
1543  * This is called from ipath_qp_rcv() to process an incoming RC packet
1544  * for the given QP.
1545  * Called at interrupt level.
1546  */
1547 void ipath_rc_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
1548                   int has_grh, void *data, u32 tlen, struct ipath_qp *qp)
1549 {
1550         struct ipath_other_headers *ohdr;
1551         u32 opcode;
1552         u32 hdrsize;
1553         u32 psn;
1554         u32 pad;
1555         struct ib_wc wc;
1556         u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
1557         int diff;
1558         struct ib_reth *reth;
1559         int header_in_data;
1560
1561         /* Validate the SLID. See Ch. 9.6.1.5 */
1562         if (unlikely(be16_to_cpu(hdr->lrh[3]) != qp->remote_ah_attr.dlid))
1563                 goto done;
1564
1565         /* Check for GRH */
1566         if (!has_grh) {
1567                 ohdr = &hdr->u.oth;
1568                 hdrsize = 8 + 12;       /* LRH + BTH */
1569                 psn = be32_to_cpu(ohdr->bth[2]);
1570                 header_in_data = 0;
1571         } else {
1572                 ohdr = &hdr->u.l.oth;
1573                 hdrsize = 8 + 40 + 12;  /* LRH + GRH + BTH */
1574                 /*
1575                  * The header with GRH is 60 bytes and the core driver sets
1576                  * the eager header buffer size to 56 bytes so the last 4
1577                  * bytes of the BTH header (PSN) is in the data buffer.
1578                  */
1579                 header_in_data = dev->dd->ipath_rcvhdrentsize == 16;
1580                 if (header_in_data) {
1581                         psn = be32_to_cpu(((__be32 *) data)[0]);
1582                         data += sizeof(__be32);
1583                 } else
1584                         psn = be32_to_cpu(ohdr->bth[2]);
1585         }
1586
1587         /*
1588          * Process responses (ACKs) before anything else.  Note that the
1589          * packet sequence number will be for something in the send work
1590          * queue rather than the expected receive packet sequence number.
1591          * In other words, this QP is the requester.
1592          */
1593         opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
1594         if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
1595             opcode <= OP(ATOMIC_ACKNOWLEDGE)) {
1596                 ipath_rc_rcv_resp(dev, ohdr, data, tlen, qp, opcode, psn,
1597                                   hdrsize, pmtu, header_in_data);
1598                 goto done;
1599         }
1600
1601         /* Compute 24 bits worth of difference. */
1602         diff = ipath_cmp24(psn, qp->r_psn);
1603         if (unlikely(diff)) {
1604                 if (ipath_rc_rcv_error(dev, ohdr, data, qp, opcode,
1605                                        psn, diff, header_in_data))
1606                         goto done;
1607                 goto send_ack;
1608         }
1609
1610         /* Check for opcode sequence errors. */
1611         switch (qp->r_state) {
1612         case OP(SEND_FIRST):
1613         case OP(SEND_MIDDLE):
1614                 if (opcode == OP(SEND_MIDDLE) ||
1615                     opcode == OP(SEND_LAST) ||
1616                     opcode == OP(SEND_LAST_WITH_IMMEDIATE))
1617                         break;
1618         nack_inv:
1619                 ipath_rc_error(qp, IB_WC_REM_INV_REQ_ERR);
1620                 qp->r_nak_state = IB_NAK_INVALID_REQUEST;
1621                 qp->r_ack_psn = qp->r_psn;
1622                 goto send_ack;
1623
1624         case OP(RDMA_WRITE_FIRST):
1625         case OP(RDMA_WRITE_MIDDLE):
1626                 if (opcode == OP(RDMA_WRITE_MIDDLE) ||
1627                     opcode == OP(RDMA_WRITE_LAST) ||
1628                     opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
1629                         break;
1630                 goto nack_inv;
1631
1632         default:
1633                 if (opcode == OP(SEND_MIDDLE) ||
1634                     opcode == OP(SEND_LAST) ||
1635                     opcode == OP(SEND_LAST_WITH_IMMEDIATE) ||
1636                     opcode == OP(RDMA_WRITE_MIDDLE) ||
1637                     opcode == OP(RDMA_WRITE_LAST) ||
1638                     opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
1639                         goto nack_inv;
1640                 /*
1641                  * Note that it is up to the requester to not send a new
1642                  * RDMA read or atomic operation before receiving an ACK
1643                  * for the previous operation.
1644                  */
1645                 break;
1646         }
1647
1648         wc.imm_data = 0;
1649         wc.wc_flags = 0;
1650
1651         /* OK, process the packet. */
1652         switch (opcode) {
1653         case OP(SEND_FIRST):
1654                 if (!ipath_get_rwqe(qp, 0)) {
1655                 rnr_nak:
1656                         /*
1657                          * A RNR NAK will ACK earlier sends and RDMA writes.
1658                          * Don't queue the NAK if a RDMA read or atomic
1659                          * is pending though.
1660                          */
1661                         if (qp->r_nak_state)
1662                                 goto done;
1663                         qp->r_nak_state = IB_RNR_NAK | qp->r_min_rnr_timer;
1664                         qp->r_ack_psn = qp->r_psn;
1665                         goto send_ack;
1666                 }
1667                 qp->r_rcv_len = 0;
1668                 /* FALLTHROUGH */
1669         case OP(SEND_MIDDLE):
1670         case OP(RDMA_WRITE_MIDDLE):
1671         send_middle:
1672                 /* Check for invalid length PMTU or posted rwqe len. */
1673                 if (unlikely(tlen != (hdrsize + pmtu + 4)))
1674                         goto nack_inv;
1675                 qp->r_rcv_len += pmtu;
1676                 if (unlikely(qp->r_rcv_len > qp->r_len))
1677                         goto nack_inv;
1678                 ipath_copy_sge(&qp->r_sge, data, pmtu);
1679                 break;
1680
1681         case OP(RDMA_WRITE_LAST_WITH_IMMEDIATE):
1682                 /* consume RWQE */
1683                 if (!ipath_get_rwqe(qp, 1))
1684                         goto rnr_nak;
1685                 goto send_last_imm;
1686
1687         case OP(SEND_ONLY):
1688         case OP(SEND_ONLY_WITH_IMMEDIATE):
1689                 if (!ipath_get_rwqe(qp, 0))
1690                         goto rnr_nak;
1691                 qp->r_rcv_len = 0;
1692                 if (opcode == OP(SEND_ONLY))
1693                         goto send_last;
1694                 /* FALLTHROUGH */
1695         case OP(SEND_LAST_WITH_IMMEDIATE):
1696         send_last_imm:
1697                 if (header_in_data) {
1698                         wc.imm_data = *(__be32 *) data;
1699                         data += sizeof(__be32);
1700                 } else {
1701                         /* Immediate data comes after BTH */
1702                         wc.imm_data = ohdr->u.imm_data;
1703                 }
1704                 hdrsize += 4;
1705                 wc.wc_flags = IB_WC_WITH_IMM;
1706                 /* FALLTHROUGH */
1707         case OP(SEND_LAST):
1708         case OP(RDMA_WRITE_LAST):
1709         send_last:
1710                 /* Get the number of bytes the message was padded by. */
1711                 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1712                 /* Check for invalid length. */
1713                 /* XXX LAST len should be >= 1 */
1714                 if (unlikely(tlen < (hdrsize + pad + 4)))
1715                         goto nack_inv;
1716                 /* Don't count the CRC. */
1717                 tlen -= (hdrsize + pad + 4);
1718                 wc.byte_len = tlen + qp->r_rcv_len;
1719                 if (unlikely(wc.byte_len > qp->r_len))
1720                         goto nack_inv;
1721                 ipath_copy_sge(&qp->r_sge, data, tlen);
1722                 qp->r_msn++;
1723                 if (!qp->r_wrid_valid)
1724                         break;
1725                 qp->r_wrid_valid = 0;
1726                 wc.wr_id = qp->r_wr_id;
1727                 wc.status = IB_WC_SUCCESS;
1728                 wc.opcode = IB_WC_RECV;
1729                 wc.vendor_err = 0;
1730                 wc.qp = &qp->ibqp;
1731                 wc.src_qp = qp->remote_qpn;
1732                 wc.pkey_index = 0;
1733                 wc.slid = qp->remote_ah_attr.dlid;
1734                 wc.sl = qp->remote_ah_attr.sl;
1735                 wc.dlid_path_bits = 0;
1736                 wc.port_num = 0;
1737                 /* Signal completion event if the solicited bit is set. */
1738                 ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
1739                                (ohdr->bth[0] &
1740                                 __constant_cpu_to_be32(1 << 23)) != 0);
1741                 break;
1742
1743         case OP(RDMA_WRITE_FIRST):
1744         case OP(RDMA_WRITE_ONLY):
1745         case OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE):
1746                 if (unlikely(!(qp->qp_access_flags &
1747                                IB_ACCESS_REMOTE_WRITE)))
1748                         goto nack_inv;
1749                 /* consume RWQE */
1750                 /* RETH comes after BTH */
1751                 if (!header_in_data)
1752                         reth = &ohdr->u.rc.reth;
1753                 else {
1754                         reth = (struct ib_reth *)data;
1755                         data += sizeof(*reth);
1756                 }
1757                 hdrsize += sizeof(*reth);
1758                 qp->r_len = be32_to_cpu(reth->length);
1759                 qp->r_rcv_len = 0;
1760                 if (qp->r_len != 0) {
1761                         u32 rkey = be32_to_cpu(reth->rkey);
1762                         u64 vaddr = be64_to_cpu(reth->vaddr);
1763                         int ok;
1764
1765                         /* Check rkey & NAK */
1766                         ok = ipath_rkey_ok(qp, &qp->r_sge,
1767                                            qp->r_len, vaddr, rkey,
1768                                            IB_ACCESS_REMOTE_WRITE);
1769                         if (unlikely(!ok))
1770                                 goto nack_acc;
1771                 } else {
1772                         qp->r_sge.sg_list = NULL;
1773                         qp->r_sge.sge.mr = NULL;
1774                         qp->r_sge.sge.vaddr = NULL;
1775                         qp->r_sge.sge.length = 0;
1776                         qp->r_sge.sge.sge_length = 0;
1777                 }
1778                 if (opcode == OP(RDMA_WRITE_FIRST))
1779                         goto send_middle;
1780                 else if (opcode == OP(RDMA_WRITE_ONLY))
1781                         goto send_last;
1782                 if (!ipath_get_rwqe(qp, 1))
1783                         goto rnr_nak;
1784                 goto send_last_imm;
1785
1786         case OP(RDMA_READ_REQUEST): {
1787                 struct ipath_ack_entry *e;
1788                 u32 len;
1789                 u8 next;
1790
1791                 if (unlikely(!(qp->qp_access_flags &
1792                                IB_ACCESS_REMOTE_READ)))
1793                         goto nack_inv;
1794                 next = qp->r_head_ack_queue + 1;
1795                 if (next > IPATH_MAX_RDMA_ATOMIC)
1796                         next = 0;
1797                 if (unlikely(next == qp->s_tail_ack_queue)) {
1798                         if (!qp->s_ack_queue[next].sent)
1799                                 goto nack_inv;
1800                         ipath_update_ack_queue(qp, next);
1801                 }
1802                 e = &qp->s_ack_queue[qp->r_head_ack_queue];
1803                 /* RETH comes after BTH */
1804                 if (!header_in_data)
1805                         reth = &ohdr->u.rc.reth;
1806                 else {
1807                         reth = (struct ib_reth *)data;
1808                         data += sizeof(*reth);
1809                 }
1810                 len = be32_to_cpu(reth->length);
1811                 if (len) {
1812                         u32 rkey = be32_to_cpu(reth->rkey);
1813                         u64 vaddr = be64_to_cpu(reth->vaddr);
1814                         int ok;
1815
1816                         /* Check rkey & NAK */
1817                         ok = ipath_rkey_ok(qp, &e->rdma_sge, len, vaddr,
1818                                            rkey, IB_ACCESS_REMOTE_READ);
1819                         if (unlikely(!ok))
1820                                 goto nack_acc;
1821                         /*
1822                          * Update the next expected PSN.  We add 1 later
1823                          * below, so only add the remainder here.
1824                          */
1825                         if (len > pmtu)
1826                                 qp->r_psn += (len - 1) / pmtu;
1827                 } else {
1828                         e->rdma_sge.sg_list = NULL;
1829                         e->rdma_sge.num_sge = 0;
1830                         e->rdma_sge.sge.mr = NULL;
1831                         e->rdma_sge.sge.vaddr = NULL;
1832                         e->rdma_sge.sge.length = 0;
1833                         e->rdma_sge.sge.sge_length = 0;
1834                 }
1835                 e->opcode = opcode;
1836                 e->sent = 0;
1837                 e->psn = psn;
1838                 /*
1839                  * We need to increment the MSN here instead of when we
1840                  * finish sending the result since a duplicate request would
1841                  * increment it more than once.
1842                  */
1843                 qp->r_msn++;
1844                 qp->r_psn++;
1845                 qp->r_state = opcode;
1846                 qp->r_nak_state = 0;
1847                 barrier();
1848                 qp->r_head_ack_queue = next;
1849
1850                 /* Call ipath_do_rc_send() in another thread. */
1851                 tasklet_hi_schedule(&qp->s_task);
1852
1853                 goto done;
1854         }
1855
1856         case OP(COMPARE_SWAP):
1857         case OP(FETCH_ADD): {
1858                 struct ib_atomic_eth *ateth;
1859                 struct ipath_ack_entry *e;
1860                 u64 vaddr;
1861                 atomic64_t *maddr;
1862                 u64 sdata;
1863                 u32 rkey;
1864                 u8 next;
1865
1866                 if (unlikely(!(qp->qp_access_flags &
1867                                IB_ACCESS_REMOTE_ATOMIC)))
1868                         goto nack_inv;
1869                 next = qp->r_head_ack_queue + 1;
1870                 if (next > IPATH_MAX_RDMA_ATOMIC)
1871                         next = 0;
1872                 if (unlikely(next == qp->s_tail_ack_queue)) {
1873                         if (!qp->s_ack_queue[next].sent)
1874                                 goto nack_inv;
1875                         ipath_update_ack_queue(qp, next);
1876                 }
1877                 if (!header_in_data)
1878                         ateth = &ohdr->u.atomic_eth;
1879                 else
1880                         ateth = (struct ib_atomic_eth *)data;
1881                 vaddr = ((u64) be32_to_cpu(ateth->vaddr[0]) << 32) |
1882                         be32_to_cpu(ateth->vaddr[1]);
1883                 if (unlikely(vaddr & (sizeof(u64) - 1)))
1884                         goto nack_inv;
1885                 rkey = be32_to_cpu(ateth->rkey);
1886                 /* Check rkey & NAK */
1887                 if (unlikely(!ipath_rkey_ok(qp, &qp->r_sge,
1888                                             sizeof(u64), vaddr, rkey,
1889                                             IB_ACCESS_REMOTE_ATOMIC)))
1890                         goto nack_acc;
1891                 /* Perform atomic OP and save result. */
1892                 maddr = (atomic64_t *) qp->r_sge.sge.vaddr;
1893                 sdata = be64_to_cpu(ateth->swap_data);
1894                 e = &qp->s_ack_queue[qp->r_head_ack_queue];
1895                 e->atomic_data = (opcode == OP(FETCH_ADD)) ?
1896                         (u64) atomic64_add_return(sdata, maddr) - sdata :
1897                         (u64) cmpxchg((u64 *) qp->r_sge.sge.vaddr,
1898                                       be64_to_cpu(ateth->compare_data),
1899                                       sdata);
1900                 e->opcode = opcode;
1901                 e->sent = 0;
1902                 e->psn = psn & IPATH_PSN_MASK;
1903                 qp->r_msn++;
1904                 qp->r_psn++;
1905                 qp->r_state = opcode;
1906                 qp->r_nak_state = 0;
1907                 barrier();
1908                 qp->r_head_ack_queue = next;
1909
1910                 /* Call ipath_do_rc_send() in another thread. */
1911                 tasklet_hi_schedule(&qp->s_task);
1912
1913                 goto done;
1914         }
1915
1916         default:
1917                 /* NAK unknown opcodes. */
1918                 goto nack_inv;
1919         }
1920         qp->r_psn++;
1921         qp->r_state = opcode;
1922         qp->r_ack_psn = psn;
1923         qp->r_nak_state = 0;
1924         /* Send an ACK if requested or required. */
1925         if (psn & (1 << 31))
1926                 goto send_ack;
1927         goto done;
1928
1929 nack_acc:
1930         ipath_rc_error(qp, IB_WC_REM_ACCESS_ERR);
1931         qp->r_nak_state = IB_NAK_REMOTE_ACCESS_ERROR;
1932         qp->r_ack_psn = qp->r_psn;
1933
1934 send_ack:
1935         send_rc_ack(qp);
1936
1937 done:
1938         return;
1939 }