drivers/infiniband/hw/ipath/ipath_rc.c

   1 /*
   2  * Copyright (c) 2006, 2007, 2008 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 <linux/io.h>
  35
  36 #include "ipath_verbs.h"
  37 #include "ipath_kernel.h"
  38
  39 /* cut down ridiculously long IB macro names */
  40 #define OP(x) IB_OPCODE_RC_##x
  41
  42 static u32 restart_sge(struct ipath_sge_state *ss, struct ipath_swqe *wqe,
  43                        u32 psn, u32 pmtu)
  44 {
  45         u32 len;
  46
  47         len = ((psn - wqe->psn) & IPATH_PSN_MASK) * pmtu;
  48         ss->sge = wqe->sg_list[0];
  49         ss->sg_list = wqe->sg_list + 1;
  50         ss->num_sge = wqe->wr.num_sge;
  51         ipath_skip_sge(ss, len);
  52         return wqe->length - len;
  53 }
  54
  55 /**
  56  * ipath_init_restart- initialize the qp->s_sge after a restart
  57  * @qp: the QP who's SGE we're restarting
  58  * @wqe: the work queue to initialize the QP's SGE from
  59  *
  60  * The QP s_lock should be held and interrupts disabled.
  61  */
  62 static void ipath_init_restart(struct ipath_qp *qp, struct ipath_swqe *wqe)
  63 {
  64         struct ipath_ibdev *dev;
  65
  66         qp->s_len = restart_sge(&qp->s_sge, wqe, qp->s_psn,
  67                                 ib_mtu_enum_to_int(qp->path_mtu));
  68         dev = to_idev(qp->ibqp.device);
  69         spin_lock(&dev->pending_lock);
  70         if (list_empty(&qp->timerwait))
  71                 list_add_tail(&qp->timerwait,
  72                               &dev->pending[dev->pending_index]);
  73         spin_unlock(&dev->pending_lock);
  74 }
  75
  76 /**
  77  * ipath_make_rc_ack - construct a response packet (ACK, NAK, or RDMA read)
  78  * @qp: a pointer to the QP
  79  * @ohdr: a pointer to the IB header being constructed
  80  * @pmtu: the path MTU
  81  *
  82  * Return 1 if constructed; otherwise, return 0.
  83  * Note that we are in the responder's side of the QP context.
  84  * Note the QP s_lock must be held.
  85  */
  86 static int ipath_make_rc_ack(struct ipath_ibdev *dev, struct ipath_qp *qp,
  87                              struct ipath_other_headers *ohdr, u32 pmtu)
  88 {
  89         struct ipath_ack_entry *e;
  90         u32 hwords;
  91         u32 len;
  92         u32 bth0;
  93         u32 bth2;
  94
  95         /* header size in 32-bit words LRH+BTH = (8+12)/4. */
  96         hwords = 5;
  97
  98         switch (qp->s_ack_state) {
  99         case OP(RDMA_READ_RESPONSE_LAST):
 100         case OP(RDMA_READ_RESPONSE_ONLY):
 101         case OP(ATOMIC_ACKNOWLEDGE):
 102                 /*
 103                  * We can increment the tail pointer now that the last
 104                  * response has been sent instead of only being
 105                  * constructed.
 106                  */
 107                 if (++qp->s_tail_ack_queue > IPATH_MAX_RDMA_ATOMIC)
 108                         qp->s_tail_ack_queue = 0;
 109                 /* FALLTHROUGH */
 110         case OP(SEND_ONLY):
 111         case OP(ACKNOWLEDGE):
 112                 /* Check for no next entry in the queue. */
 113                 if (qp->r_head_ack_queue == qp->s_tail_ack_queue) {
 114                         if (qp->s_flags & IPATH_S_ACK_PENDING)
 115                                 goto normal;
 116                         qp->s_ack_state = OP(ACKNOWLEDGE);
 117                         goto bail;
 118                 }
 119
 120                 e = &qp->s_ack_queue[qp->s_tail_ack_queue];
 121                 if (e->opcode == OP(RDMA_READ_REQUEST)) {
 122                         /* Copy SGE state in case we need to resend */
 123                         qp->s_ack_rdma_sge = e->rdma_sge;
 124                         qp->s_cur_sge = &qp->s_ack_rdma_sge;
 125                         len = e->rdma_sge.sge.sge_length;
 126                         if (len > pmtu) {
 127                                 len = pmtu;
 128                                 qp->s_ack_state = OP(RDMA_READ_RESPONSE_FIRST);
 129                         } else {
 130                                 qp->s_ack_state = OP(RDMA_READ_RESPONSE_ONLY);
 131                                 e->sent = 1;
 132                         }
 133                         ohdr->u.aeth = ipath_compute_aeth(qp);
 134                         hwords++;
 135                         qp->s_ack_rdma_psn = e->psn;
 136                         bth2 = qp->s_ack_rdma_psn++ & IPATH_PSN_MASK;
 137                 } else {
 138                         /* COMPARE_SWAP or FETCH_ADD */
 139                         qp->s_cur_sge = NULL;
 140                         len = 0;
 141                         qp->s_ack_state = OP(ATOMIC_ACKNOWLEDGE);
 142                         ohdr->u.at.aeth = ipath_compute_aeth(qp);
 143                         ohdr->u.at.atomic_ack_eth[0] =
 144                                 cpu_to_be32(e->atomic_data >> 32);
 145                         ohdr->u.at.atomic_ack_eth[1] =
 146                                 cpu_to_be32(e->atomic_data);
 147                         hwords += sizeof(ohdr->u.at) / sizeof(u32);
 148                         bth2 = e->psn;
 149                         e->sent = 1;
 150                 }
 151                 bth0 = qp->s_ack_state << 24;
 152                 break;
 153
 154         case OP(RDMA_READ_RESPONSE_FIRST):
 155                 qp->s_ack_state = OP(RDMA_READ_RESPONSE_MIDDLE);
 156                 /* FALLTHROUGH */
 157         case OP(RDMA_READ_RESPONSE_MIDDLE):
 158                 len = qp->s_ack_rdma_sge.sge.sge_length;
 159                 if (len > pmtu)
 160                         len = pmtu;
 161                 else {
 162                         ohdr->u.aeth = ipath_compute_aeth(qp);
 163                         hwords++;
 164                         qp->s_ack_state = OP(RDMA_READ_RESPONSE_LAST);
 165                         qp->s_ack_queue[qp->s_tail_ack_queue].sent = 1;
 166                 }
 167                 bth0 = qp->s_ack_state << 24;
 168                 bth2 = qp->s_ack_rdma_psn++ & IPATH_PSN_MASK;
 169                 break;
 170
 171         default:
 172         normal:
 173                 /*
 174                  * Send a regular ACK.
 175                  * Set the s_ack_state so we wait until after sending
 176                  * the ACK before setting s_ack_state to ACKNOWLEDGE
 177                  * (see above).
 178                  */
 179                 qp->s_ack_state = OP(SEND_ONLY);
 180                 qp->s_flags &= ~IPATH_S_ACK_PENDING;
 181                 qp->s_cur_sge = NULL;
 182                 if (qp->s_nak_state)
 183                         ohdr->u.aeth =
 184                                 cpu_to_be32((qp->r_msn & IPATH_MSN_MASK) |
 185                                             (qp->s_nak_state <<
 186                                              IPATH_AETH_CREDIT_SHIFT));
 187                 else
 188                         ohdr->u.aeth = ipath_compute_aeth(qp);
 189                 hwords++;
 190                 len = 0;
 191                 bth0 = OP(ACKNOWLEDGE) << 24;
 192                 bth2 = qp->s_ack_psn & IPATH_PSN_MASK;
 193         }
 194         qp->s_hdrwords = hwords;
 195         qp->s_cur_size = len;
 196         ipath_make_ruc_header(dev, qp, ohdr, bth0, bth2);
 197         return 1;
 198
 199 bail:
 200         return 0;
 201 }
 202
 203 /**
 204  * ipath_make_rc_req - construct a request packet (SEND, RDMA r/w, ATOMIC)
 205  * @qp: a pointer to the QP
 206  *
 207  * Return 1 if constructed; otherwise, return 0.
 208  */
 209 int ipath_make_rc_req(struct ipath_qp *qp)
 210 {
 211         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
 212         struct ipath_other_headers *ohdr;
 213         struct ipath_sge_state *ss;
 214         struct ipath_swqe *wqe;
 215         u32 hwords;
 216         u32 len;
 217         u32 bth0;
 218         u32 bth2;
 219         u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
 220         char newreq;
 221         unsigned long flags;
 222         int ret = 0;
 223
 224         ohdr = &qp->s_hdr.u.oth;
 225         if (qp->remote_ah_attr.ah_flags & IB_AH_GRH)
 226                 ohdr = &qp->s_hdr.u.l.oth;
 227
 228         /*
 229          * The lock is needed to synchronize between the sending tasklet,
 230          * the receive interrupt handler, and timeout resends.
 231          */
 232         spin_lock_irqsave(&qp->s_lock, flags);
 233
 234         /* Sending responses has higher priority over sending requests. */
 235         if ((qp->r_head_ack_queue != qp->s_tail_ack_queue ||
 236              (qp->s_flags & IPATH_S_ACK_PENDING) ||
 237              qp->s_ack_state != OP(ACKNOWLEDGE)) &&
 238             ipath_make_rc_ack(dev, qp, ohdr, pmtu))
 239                 goto done;
 240
 241         if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK) ||
 242             qp->s_rnr_timeout || qp->s_wait_credit)
 243                 goto bail;
 244
 245         /* Limit the number of packets sent without an ACK. */
 246         if (ipath_cmp24(qp->s_psn, qp->s_last_psn + IPATH_PSN_CREDIT) > 0) {
 247                 qp->s_wait_credit = 1;
 248                 dev->n_rc_stalls++;
 249                 goto bail;
 250         }
 251
 252         /* header size in 32-bit words LRH+BTH = (8+12)/4. */
 253         hwords = 5;
 254         bth0 = 1 << 22; /* Set M bit */
 255
 256         /* Send a request. */
 257         wqe = get_swqe_ptr(qp, qp->s_cur);
 258         switch (qp->s_state) {
 259         default:
 260                 /*
 261                  * Resend an old request or start a new one.
 262                  *
 263                  * We keep track of the current SWQE so that
 264                  * we don't reset the "furthest progress" state
 265                  * if we need to back up.
 266                  */
 267                 newreq = 0;
 268                 if (qp->s_cur == qp->s_tail) {
 269                         /* Check if send work queue is empty. */
 270                         if (qp->s_tail == qp->s_head)
 271                                 goto bail;
 272                         /*
 273                          * If a fence is requested, wait for previous
 274                          * RDMA read and atomic operations to finish.
 275                          */
 276                         if ((wqe->wr.send_flags & IB_SEND_FENCE) &&
 277                             qp->s_num_rd_atomic) {
 278                                 qp->s_flags |= IPATH_S_FENCE_PENDING;
 279                                 goto bail;
 280                         }
 281                         wqe->psn = qp->s_next_psn;
 282                         newreq = 1;
 283                 }
 284                 /*
 285                  * Note that we have to be careful not to modify the
 286                  * original work request since we may need to resend
 287                  * it.
 288                  */
 289                 len = wqe->length;
 290                 ss = &qp->s_sge;
 291                 bth2 = 0;
 292                 switch (wqe->wr.opcode) {
 293                 case IB_WR_SEND:
 294                 case IB_WR_SEND_WITH_IMM:
 295                         /* If no credit, return. */
 296                         if (qp->s_lsn != (u32) -1 &&
 297                             ipath_cmp24(wqe->ssn, qp->s_lsn + 1) > 0)
 298                                 goto bail;
 299                         wqe->lpsn = wqe->psn;
 300                         if (len > pmtu) {
 301                                 wqe->lpsn += (len - 1) / pmtu;
 302                                 qp->s_state = OP(SEND_FIRST);
 303                                 len = pmtu;
 304                                 break;
 305                         }
 306                         if (wqe->wr.opcode == IB_WR_SEND)
 307                                 qp->s_state = OP(SEND_ONLY);
 308                         else {
 309                                 qp->s_state = OP(SEND_ONLY_WITH_IMMEDIATE);
 310                                 /* Immediate data comes after the BTH */
 311                                 ohdr->u.imm_data = wqe->wr.ex.imm_data;
 312                                 hwords += 1;
 313                         }
 314                         if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 315                                 bth0 |= 1 << 23;
 316                         bth2 = 1 << 31; /* Request ACK. */
 317                         if (++qp->s_cur == qp->s_size)
 318                                 qp->s_cur = 0;
 319                         break;
 320
 321                 case IB_WR_RDMA_WRITE:
 322                         if (newreq && qp->s_lsn != (u32) -1)
 323                                 qp->s_lsn++;
 324                         /* FALLTHROUGH */
 325                 case IB_WR_RDMA_WRITE_WITH_IMM:
 326                         /* If no credit, return. */
 327                         if (qp->s_lsn != (u32) -1 &&
 328                             ipath_cmp24(wqe->ssn, qp->s_lsn + 1) > 0)
 329                                 goto bail;
 330                         ohdr->u.rc.reth.vaddr =
 331                                 cpu_to_be64(wqe->wr.wr.rdma.remote_addr);
 332                         ohdr->u.rc.reth.rkey =
 333                                 cpu_to_be32(wqe->wr.wr.rdma.rkey);
 334                         ohdr->u.rc.reth.length = cpu_to_be32(len);
 335                         hwords += sizeof(struct ib_reth) / sizeof(u32);
 336                         wqe->lpsn = wqe->psn;
 337                         if (len > pmtu) {
 338                                 wqe->lpsn += (len - 1) / pmtu;
 339                                 qp->s_state = OP(RDMA_WRITE_FIRST);
 340                                 len = pmtu;
 341                                 break;
 342                         }
 343                         if (wqe->wr.opcode == IB_WR_RDMA_WRITE)
 344                                 qp->s_state = OP(RDMA_WRITE_ONLY);
 345                         else {
 346                                 qp->s_state =
 347                                         OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE);
 348                                 /* Immediate data comes after RETH */
 349                                 ohdr->u.rc.imm_data = wqe->wr.ex.imm_data;
 350                                 hwords += 1;
 351                                 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 352                                         bth0 |= 1 << 23;
 353                         }
 354                         bth2 = 1 << 31; /* Request ACK. */
 355                         if (++qp->s_cur == qp->s_size)
 356                                 qp->s_cur = 0;
 357                         break;
 358
 359                 case IB_WR_RDMA_READ:
 360                         /*
 361                          * Don't allow more operations to be started
 362                          * than the QP limits allow.
 363                          */
 364                         if (newreq) {
 365                                 if (qp->s_num_rd_atomic >=
 366                                     qp->s_max_rd_atomic) {
 367                                         qp->s_flags |= IPATH_S_RDMAR_PENDING;
 368                                         goto bail;
 369                                 }
 370                                 qp->s_num_rd_atomic++;
 371                                 if (qp->s_lsn != (u32) -1)
 372                                         qp->s_lsn++;
 373                                 /*
 374                                  * Adjust s_next_psn to count the
 375                                  * expected number of responses.
 376                                  */
 377                                 if (len > pmtu)
 378                                         qp->s_next_psn += (len - 1) / pmtu;
 379                                 wqe->lpsn = qp->s_next_psn++;
 380                         }
 381                         ohdr->u.rc.reth.vaddr =
 382                                 cpu_to_be64(wqe->wr.wr.rdma.remote_addr);
 383                         ohdr->u.rc.reth.rkey =
 384                                 cpu_to_be32(wqe->wr.wr.rdma.rkey);
 385                         ohdr->u.rc.reth.length = cpu_to_be32(len);
 386                         qp->s_state = OP(RDMA_READ_REQUEST);
 387                         hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
 388                         ss = NULL;
 389                         len = 0;
 390                         if (++qp->s_cur == qp->s_size)
 391                                 qp->s_cur = 0;
 392                         break;
 393
 394                 case IB_WR_ATOMIC_CMP_AND_SWP:
 395                 case IB_WR_ATOMIC_FETCH_AND_ADD:
 396                         /*
 397                          * Don't allow more operations to be started
 398                          * than the QP limits allow.
 399                          */
 400                         if (newreq) {
 401                                 if (qp->s_num_rd_atomic >=
 402                                     qp->s_max_rd_atomic) {
 403                                         qp->s_flags |= IPATH_S_RDMAR_PENDING;
 404                                         goto bail;
 405                                 }
 406                                 qp->s_num_rd_atomic++;
 407                                 if (qp->s_lsn != (u32) -1)
 408                                         qp->s_lsn++;
 409                                 wqe->lpsn = wqe->psn;
 410                         }
 411                         if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
 412                                 qp->s_state = OP(COMPARE_SWAP);
 413                                 ohdr->u.atomic_eth.swap_data = cpu_to_be64(
 414                                         wqe->wr.wr.atomic.swap);
 415                                 ohdr->u.atomic_eth.compare_data = cpu_to_be64(
 416                                         wqe->wr.wr.atomic.compare_add);
 417                         } else {
 418                                 qp->s_state = OP(FETCH_ADD);
 419                                 ohdr->u.atomic_eth.swap_data = cpu_to_be64(
 420                                         wqe->wr.wr.atomic.compare_add);
 421                                 ohdr->u.atomic_eth.compare_data = 0;
 422                         }
 423                         ohdr->u.atomic_eth.vaddr[0] = cpu_to_be32(
 424                                 wqe->wr.wr.atomic.remote_addr >> 32);
 425                         ohdr->u.atomic_eth.vaddr[1] = cpu_to_be32(
 426                                 wqe->wr.wr.atomic.remote_addr);
 427                         ohdr->u.atomic_eth.rkey = cpu_to_be32(
 428                                 wqe->wr.wr.atomic.rkey);
 429                         hwords += sizeof(struct ib_atomic_eth) / sizeof(u32);
 430                         ss = NULL;
 431                         len = 0;
 432                         if (++qp->s_cur == qp->s_size)
 433                                 qp->s_cur = 0;
 434                         break;
 435
 436                 default:
 437                         goto bail;
 438                 }
 439                 qp->s_sge.sge = wqe->sg_list[0];
 440                 qp->s_sge.sg_list = wqe->sg_list + 1;
 441                 qp->s_sge.num_sge = wqe->wr.num_sge;
 442                 qp->s_len = wqe->length;
 443                 if (newreq) {
 444                         qp->s_tail++;
 445                         if (qp->s_tail >= qp->s_size)
 446                                 qp->s_tail = 0;
 447                 }
 448                 bth2 |= qp->s_psn & IPATH_PSN_MASK;
 449                 if (wqe->wr.opcode == IB_WR_RDMA_READ)
 450                         qp->s_psn = wqe->lpsn + 1;
 451                 else {
 452                         qp->s_psn++;
 453                         if (ipath_cmp24(qp->s_psn, qp->s_next_psn) > 0)
 454                                 qp->s_next_psn = qp->s_psn;
 455                 }
 456                 /*
 457                  * Put the QP on the pending list so lost ACKs will cause
 458                  * a retry.  More than one request can be pending so the
 459                  * QP may already be on the dev->pending list.
 460                  */
 461                 spin_lock(&dev->pending_lock);
 462                 if (list_empty(&qp->timerwait))
 463                         list_add_tail(&qp->timerwait,
 464                                       &dev->pending[dev->pending_index]);
 465                 spin_unlock(&dev->pending_lock);
 466                 break;
 467
 468         case OP(RDMA_READ_RESPONSE_FIRST):
 469                 /*
 470                  * This case can only happen if a send is restarted.
 471                  * See ipath_restart_rc().
 472                  */
 473                 ipath_init_restart(qp, wqe);
 474                 /* FALLTHROUGH */
 475         case OP(SEND_FIRST):
 476                 qp->s_state = OP(SEND_MIDDLE);
 477                 /* FALLTHROUGH */
 478         case OP(SEND_MIDDLE):
 479                 bth2 = qp->s_psn++ & IPATH_PSN_MASK;
 480                 if (ipath_cmp24(qp->s_psn, qp->s_next_psn) > 0)
 481                         qp->s_next_psn = qp->s_psn;
 482                 ss = &qp->s_sge;
 483                 len = qp->s_len;
 484                 if (len > pmtu) {
 485                         len = pmtu;
 486                         break;
 487                 }
 488                 if (wqe->wr.opcode == IB_WR_SEND)
 489                         qp->s_state = OP(SEND_LAST);
 490                 else {
 491                         qp->s_state = OP(SEND_LAST_WITH_IMMEDIATE);
 492                         /* Immediate data comes after the BTH */
 493                         ohdr->u.imm_data = wqe->wr.ex.imm_data;
 494                         hwords += 1;
 495                 }
 496                 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 497                         bth0 |= 1 << 23;
 498                 bth2 |= 1 << 31;        /* Request ACK. */
 499                 qp->s_cur++;
 500                 if (qp->s_cur >= qp->s_size)
 501                         qp->s_cur = 0;
 502                 break;
 503
 504         case OP(RDMA_READ_RESPONSE_LAST):
 505                 /*
 506                  * This case can only happen if a RDMA write is restarted.
 507                  * See ipath_restart_rc().
 508                  */
 509                 ipath_init_restart(qp, wqe);
 510                 /* FALLTHROUGH */
 511         case OP(RDMA_WRITE_FIRST):
 512                 qp->s_state = OP(RDMA_WRITE_MIDDLE);
 513                 /* FALLTHROUGH */
 514         case OP(RDMA_WRITE_MIDDLE):
 515                 bth2 = qp->s_psn++ & IPATH_PSN_MASK;
 516                 if (ipath_cmp24(qp->s_psn, qp->s_next_psn) > 0)
 517                         qp->s_next_psn = qp->s_psn;
 518                 ss = &qp->s_sge;
 519                 len = qp->s_len;
 520                 if (len > pmtu) {
 521                         len = pmtu;
 522                         break;
 523                 }
 524                 if (wqe->wr.opcode == IB_WR_RDMA_WRITE)
 525                         qp->s_state = OP(RDMA_WRITE_LAST);
 526                 else {
 527                         qp->s_state = OP(RDMA_WRITE_LAST_WITH_IMMEDIATE);
 528                         /* Immediate data comes after the BTH */
 529                         ohdr->u.imm_data = wqe->wr.ex.imm_data;
 530                         hwords += 1;
 531                         if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 532                                 bth0 |= 1 << 23;
 533                 }
 534                 bth2 |= 1 << 31;        /* Request ACK. */
 535                 qp->s_cur++;
 536                 if (qp->s_cur >= qp->s_size)
 537                         qp->s_cur = 0;
 538                 break;
 539
 540         case OP(RDMA_READ_RESPONSE_MIDDLE):
 541                 /*
 542                  * This case can only happen if a RDMA read is restarted.
 543                  * See ipath_restart_rc().
 544                  */
 545                 ipath_init_restart(qp, wqe);
 546                 len = ((qp->s_psn - wqe->psn) & IPATH_PSN_MASK) * pmtu;
 547                 ohdr->u.rc.reth.vaddr =
 548                         cpu_to_be64(wqe->wr.wr.rdma.remote_addr + len);
 549                 ohdr->u.rc.reth.rkey =
 550                         cpu_to_be32(wqe->wr.wr.rdma.rkey);
 551                 ohdr->u.rc.reth.length = cpu_to_be32(qp->s_len);
 552                 qp->s_state = OP(RDMA_READ_REQUEST);
 553                 hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
 554                 bth2 = qp->s_psn++ & IPATH_PSN_MASK;
 555                 if (ipath_cmp24(qp->s_psn, qp->s_next_psn) > 0)
 556                         qp->s_next_psn = qp->s_psn;
 557                 ss = NULL;
 558                 len = 0;
 559                 qp->s_cur++;
 560                 if (qp->s_cur == qp->s_size)
 561                         qp->s_cur = 0;
 562                 break;
 563         }
 564         if (ipath_cmp24(qp->s_psn, qp->s_last_psn + IPATH_PSN_CREDIT - 1) >= 0)
 565                 bth2 |= 1 << 31;        /* Request ACK. */
 566         qp->s_len -= len;
 567         qp->s_hdrwords = hwords;
 568         qp->s_cur_sge = ss;
 569         qp->s_cur_size = len;
 570         ipath_make_ruc_header(dev, qp, ohdr, bth0 | (qp->s_state << 24), bth2);
 571 done:
 572         ret = 1;
 573 bail:
 574         spin_unlock_irqrestore(&qp->s_lock, flags);
 575         return ret;
 576 }
 577
 578 /**
 579  * send_rc_ack - Construct an ACK packet and send it
 580  * @qp: a pointer to the QP
 581  *
 582  * This is called from ipath_rc_rcv() and only uses the receive
 583  * side QP state.
 584  * Note that RDMA reads and atomics are handled in the
 585  * send side QP state and tasklet.
 586  */
 587 static void send_rc_ack(struct ipath_qp *qp)
 588 {
 589         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
 590         struct ipath_devdata *dd;
 591         u16 lrh0;
 592         u32 bth0;
 593         u32 hwords;
 594         u32 __iomem *piobuf;
 595         struct ipath_ib_header hdr;
 596         struct ipath_other_headers *ohdr;
 597         unsigned long flags;
 598
 599         spin_lock_irqsave(&qp->s_lock, flags);
 600
 601         /* Don't send ACK or NAK if a RDMA read or atomic is pending. */
 602         if (qp->r_head_ack_queue != qp->s_tail_ack_queue ||
 603             (qp->s_flags & IPATH_S_ACK_PENDING) ||
 604             qp->s_ack_state != OP(ACKNOWLEDGE))
 605                 goto queue_ack;
 606
 607         spin_unlock_irqrestore(&qp->s_lock, flags);
 608
 609         dd = dev->dd;
 610         piobuf = ipath_getpiobuf(dd, 0, NULL);
 611         if (!piobuf) {
 612                 /*
 613                  * We are out of PIO buffers at the moment.
 614                  * Pass responsibility for sending the ACK to the
 615                  * send tasklet so that when a PIO buffer becomes
 616                  * available, the ACK is sent ahead of other outgoing
 617                  * packets.
 618                  */
 619                 spin_lock_irqsave(&qp->s_lock, flags);
 620                 goto queue_ack;
 621         }
 622
 623         /* Construct the header. */
 624         ohdr = &hdr.u.oth;
 625         lrh0 = IPATH_LRH_BTH;
 626         /* header size in 32-bit words LRH+BTH+AETH = (8+12+4)/4. */
 627         hwords = 6;
 628         if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
 629                 hwords += ipath_make_grh(dev, &hdr.u.l.grh,
 630                                          &qp->remote_ah_attr.grh,
 631                                          hwords, 0);
 632                 ohdr = &hdr.u.l.oth;
 633                 lrh0 = IPATH_LRH_GRH;
 634         }
 635         /* read pkey_index w/o lock (its atomic) */
 636         bth0 = ipath_get_pkey(dd, qp->s_pkey_index) |
 637                 (OP(ACKNOWLEDGE) << 24) | (1 << 22);
 638         if (qp->r_nak_state)
 639                 ohdr->u.aeth = cpu_to_be32((qp->r_msn & IPATH_MSN_MASK) |
 640                                             (qp->r_nak_state <<
 641                                              IPATH_AETH_CREDIT_SHIFT));
 642         else
 643                 ohdr->u.aeth = ipath_compute_aeth(qp);
 644         lrh0 |= qp->remote_ah_attr.sl << 4;
 645         hdr.lrh[0] = cpu_to_be16(lrh0);
 646         hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
 647         hdr.lrh[2] = cpu_to_be16(hwords + SIZE_OF_CRC);
 648         hdr.lrh[3] = cpu_to_be16(dd->ipath_lid);
 649         ohdr->bth[0] = cpu_to_be32(bth0);
 650         ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
 651         ohdr->bth[2] = cpu_to_be32(qp->r_ack_psn & IPATH_PSN_MASK);
 652
 653         writeq(hwords + 1, piobuf);
 654
 655         if (dd->ipath_flags & IPATH_PIO_FLUSH_WC) {
 656                 u32 *hdrp = (u32 *) &hdr;
 657
 658                 ipath_flush_wc();
 659                 __iowrite32_copy(piobuf + 2, hdrp, hwords - 1);
 660                 ipath_flush_wc();
 661                 __raw_writel(hdrp[hwords - 1], piobuf + hwords + 1);
 662         } else
 663                 __iowrite32_copy(piobuf + 2, (u32 *) &hdr, hwords);
 664
 665         ipath_flush_wc();
 666
 667         dev->n_unicast_xmit++;
 668         goto done;
 669
 670 queue_ack:
 671         dev->n_rc_qacks++;
 672         qp->s_flags |= IPATH_S_ACK_PENDING;
 673         qp->s_nak_state = qp->r_nak_state;
 674         qp->s_ack_psn = qp->r_ack_psn;
 675         spin_unlock_irqrestore(&qp->s_lock, flags);
 676
 677         /* Call ipath_do_rc_send() in another thread. */
 678         tasklet_hi_schedule(&qp->s_task);
 679
 680 done:
 681         return;
 682 }
 683
 684 /**
 685  * reset_psn - reset the QP state to send starting from PSN
 686  * @qp: the QP
 687  * @psn: the packet sequence number to restart at
 688  *
 689  * This is called from ipath_rc_rcv() to process an incoming RC ACK
 690  * for the given QP.
 691  * Called at interrupt level with the QP s_lock held.
 692  */
 693 static void reset_psn(struct ipath_qp *qp, u32 psn)
 694 {
 695         u32 n = qp->s_last;
 696         struct ipath_swqe *wqe = get_swqe_ptr(qp, n);
 697         u32 opcode;
 698
 699         qp->s_cur = n;
 700
 701         /*
 702          * If we are starting the request from the beginning,
 703          * let the normal send code handle initialization.
 704          */
 705         if (ipath_cmp24(psn, wqe->psn) <= 0) {
 706                 qp->s_state = OP(SEND_LAST);
 707                 goto done;
 708         }
 709
 710         /* Find the work request opcode corresponding to the given PSN. */
 711         opcode = wqe->wr.opcode;
 712         for (;;) {
 713                 int diff;
 714
 715                 if (++n == qp->s_size)
 716                         n = 0;
 717                 if (n == qp->s_tail)
 718                         break;
 719                 wqe = get_swqe_ptr(qp, n);
 720                 diff = ipath_cmp24(psn, wqe->psn);
 721                 if (diff < 0)
 722                         break;
 723                 qp->s_cur = n;
 724                 /*
 725                  * If we are starting the request from the beginning,
 726                  * let the normal send code handle initialization.
 727                  */
 728                 if (diff == 0) {
 729                         qp->s_state = OP(SEND_LAST);
 730                         goto done;
 731                 }
 732                 opcode = wqe->wr.opcode;
 733         }
 734
 735         /*
 736          * Set the state to restart in the middle of a request.
 737          * Don't change the s_sge, s_cur_sge, or s_cur_size.
 738          * See ipath_do_rc_send().
 739          */
 740         switch (opcode) {
 741         case IB_WR_SEND:
 742         case IB_WR_SEND_WITH_IMM:
 743                 qp->s_state = OP(RDMA_READ_RESPONSE_FIRST);
 744                 break;
 745
 746         case IB_WR_RDMA_WRITE:
 747         case IB_WR_RDMA_WRITE_WITH_IMM:
 748                 qp->s_state = OP(RDMA_READ_RESPONSE_LAST);
 749                 break;
 750
 751         case IB_WR_RDMA_READ:
 752                 qp->s_state = OP(RDMA_READ_RESPONSE_MIDDLE);
 753                 break;
 754
 755         default:
 756                 /*
 757                  * This case shouldn't happen since its only
 758                  * one PSN per req.
 759                  */
 760                 qp->s_state = OP(SEND_LAST);
 761         }
 762 done:
 763         qp->s_psn = psn;
 764 }
 765
 766 /**
 767  * ipath_restart_rc - back up requester to resend the last un-ACKed request
 768  * @qp: the QP to restart
 769  * @psn: packet sequence number for the request
 770  * @wc: the work completion request
 771  *
 772  * The QP s_lock should be held and interrupts disabled.
 773  */
 774 void ipath_restart_rc(struct ipath_qp *qp, u32 psn)
 775 {
 776         struct ipath_swqe *wqe = get_swqe_ptr(qp, qp->s_last);
 777         struct ipath_ibdev *dev;
 778
 779         if (qp->s_retry == 0) {
 780                 ipath_send_complete(qp, wqe, IB_WC_RETRY_EXC_ERR);
 781                 ipath_error_qp(qp, IB_WC_WR_FLUSH_ERR);
 782                 goto bail;
 783         }
 784         qp->s_retry--;
 785
 786         /*
 787          * Remove the QP from the timeout queue.
 788          * Note: it may already have been removed by ipath_ib_timer().
 789          */
 790         dev = to_idev(qp->ibqp.device);
 791         spin_lock(&dev->pending_lock);
 792         if (!list_empty(&qp->timerwait))
 793                 list_del_init(&qp->timerwait);
 794         if (!list_empty(&qp->piowait))
 795                 list_del_init(&qp->piowait);
 796         spin_unlock(&dev->pending_lock);
 797
 798         if (wqe->wr.opcode == IB_WR_RDMA_READ)
 799                 dev->n_rc_resends++;
 800         else
 801                 dev->n_rc_resends += (qp->s_psn - psn) & IPATH_PSN_MASK;
 802
 803         reset_psn(qp, psn);
 804         tasklet_hi_schedule(&qp->s_task);
 805
 806 bail:
 807         return;
 808 }
 809
 810 static inline void update_last_psn(struct ipath_qp *qp, u32 psn)
 811 {
 812         if (qp->s_last_psn != psn) {
 813                 qp->s_last_psn = psn;
 814                 if (qp->s_wait_credit) {
 815                         qp->s_wait_credit = 0;
 816                         tasklet_hi_schedule(&qp->s_task);
 817                 }
 818         }
 819 }
 820
 821 /**
 822  * do_rc_ack - process an incoming RC ACK
 823  * @qp: the QP the ACK came in on
 824  * @psn: the packet sequence number of the ACK
 825  * @opcode: the opcode of the request that resulted in the ACK
 826  *
 827  * This is called from ipath_rc_rcv_resp() to process an incoming RC ACK
 828  * for the given QP.
 829  * Called at interrupt level with the QP s_lock held and interrupts disabled.
 830  * Returns 1 if OK, 0 if current operation should be aborted (NAK).
 831  */
 832 static int do_rc_ack(struct ipath_qp *qp, u32 aeth, u32 psn, int opcode,
 833                      u64 val)
 834 {
 835         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
 836         struct ib_wc wc;
 837         enum ib_wc_status status;
 838         struct ipath_swqe *wqe;
 839         int ret = 0;
 840         u32 ack_psn;
 841         int diff;
 842
 843         /*
 844          * Remove the QP from the timeout queue (or RNR timeout queue).
 845          * If ipath_ib_timer() has already removed it,
 846          * it's OK since we hold the QP s_lock and ipath_restart_rc()
 847          * just won't find anything to restart if we ACK everything.
 848          */
 849         spin_lock(&dev->pending_lock);
 850         if (!list_empty(&qp->timerwait))
 851                 list_del_init(&qp->timerwait);
 852         spin_unlock(&dev->pending_lock);
 853
 854         /*
 855          * Note that NAKs implicitly ACK outstanding SEND and RDMA write
 856          * requests and implicitly NAK RDMA read and atomic requests issued
 857          * before the NAK'ed request.  The MSN won't include the NAK'ed
 858          * request but will include an ACK'ed request(s).
 859          */
 860         ack_psn = psn;
 861         if (aeth >> 29)
 862                 ack_psn--;
 863         wqe = get_swqe_ptr(qp, qp->s_last);
 864
 865         /*
 866          * The MSN might be for a later WQE than the PSN indicates so
 867          * only complete WQEs that the PSN finishes.
 868          */
 869         while ((diff = ipath_cmp24(ack_psn, wqe->lpsn)) >= 0) {
 870                 /*
 871                  * RDMA_READ_RESPONSE_ONLY is a special case since
 872                  * we want to generate completion events for everything
 873                  * before the RDMA read, copy the data, then generate
 874                  * the completion for the read.
 875                  */
 876                 if (wqe->wr.opcode == IB_WR_RDMA_READ &&
 877                     opcode == OP(RDMA_READ_RESPONSE_ONLY) &&
 878                     diff == 0) {
 879                         ret = 1;
 880                         goto bail;
 881                 }
 882                 /*
 883                  * If this request is a RDMA read or atomic, and the ACK is
 884                  * for a later operation, this ACK NAKs the RDMA read or
 885                  * atomic.  In other words, only a RDMA_READ_LAST or ONLY
 886                  * can ACK a RDMA read and likewise for atomic ops.  Note
 887                  * that the NAK case can only happen if relaxed ordering is
 888                  * used and requests are sent after an RDMA read or atomic
 889                  * is sent but before the response is received.
 890                  */
 891                 if ((wqe->wr.opcode == IB_WR_RDMA_READ &&
 892                      (opcode != OP(RDMA_READ_RESPONSE_LAST) || diff != 0)) ||
 893                     ((wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
 894                       wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) &&
 895                      (opcode != OP(ATOMIC_ACKNOWLEDGE) || diff != 0))) {
 896                         /*
 897                          * The last valid PSN seen is the previous
 898                          * request's.
 899                          */
 900                         update_last_psn(qp, wqe->psn - 1);
 901                         /* Retry this request. */
 902                         ipath_restart_rc(qp, wqe->psn);
 903                         /*
 904                          * No need to process the ACK/NAK since we are
 905                          * restarting an earlier request.
 906                          */
 907                         goto bail;
 908                 }
 909                 if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
 910                     wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)
 911                         *(u64 *) wqe->sg_list[0].vaddr = val;
 912                 if (qp->s_num_rd_atomic &&
 913                     (wqe->wr.opcode == IB_WR_RDMA_READ ||
 914                      wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
 915                      wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)) {
 916                         qp->s_num_rd_atomic--;
 917                         /* Restart sending task if fence is complete */
 918                         if ((qp->s_flags & IPATH_S_FENCE_PENDING) &&
 919                             !qp->s_num_rd_atomic) {
 920                                 qp->s_flags &= ~IPATH_S_FENCE_PENDING;
 921                                 tasklet_hi_schedule(&qp->s_task);
 922                         } else if (qp->s_flags & IPATH_S_RDMAR_PENDING) {
 923                                 qp->s_flags &= ~IPATH_S_RDMAR_PENDING;
 924                                 tasklet_hi_schedule(&qp->s_task);
 925                         }
 926                 }
 927                 /* Post a send completion queue entry if requested. */
 928                 if (!(qp->s_flags & IPATH_S_SIGNAL_REQ_WR) ||
 929                     (wqe->wr.send_flags & IB_SEND_SIGNALED)) {
 930                         memset(&wc, 0, sizeof wc);
 931                         wc.wr_id = wqe->wr.wr_id;
 932                         wc.status = IB_WC_SUCCESS;
 933                         wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
 934                         wc.byte_len = wqe->length;
 935                         wc.qp = &qp->ibqp;
 936                         wc.src_qp = qp->remote_qpn;
 937                         wc.slid = qp->remote_ah_attr.dlid;
 938                         wc.sl = qp->remote_ah_attr.sl;
 939                         ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 0);
 940                 }
 941                 qp->s_retry = qp->s_retry_cnt;
 942                 /*
 943                  * If we are completing a request which is in the process of
 944                  * being resent, we can stop resending it since we know the
 945                  * responder has already seen it.
 946                  */
 947                 if (qp->s_last == qp->s_cur) {
 948                         if (++qp->s_cur >= qp->s_size)
 949                                 qp->s_cur = 0;
 950                         qp->s_last = qp->s_cur;
 951                         if (qp->s_last == qp->s_tail)
 952                                 break;
 953                         wqe = get_swqe_ptr(qp, qp->s_cur);
 954                         qp->s_state = OP(SEND_LAST);
 955                         qp->s_psn = wqe->psn;
 956                 } else {
 957                         if (++qp->s_last >= qp->s_size)
 958                                 qp->s_last = 0;
 959                         if (qp->s_last == qp->s_tail)
 960                                 break;
 961                         wqe = get_swqe_ptr(qp, qp->s_last);
 962                 }
 963         }
 964
 965         switch (aeth >> 29) {
 966         case 0:         /* ACK */
 967                 dev->n_rc_acks++;
 968                 /* If this is a partial ACK, reset the retransmit timer. */
 969                 if (qp->s_last != qp->s_tail) {
 970                         spin_lock(&dev->pending_lock);
 971                         if (list_empty(&qp->timerwait))
 972                                 list_add_tail(&qp->timerwait,
 973                                         &dev->pending[dev->pending_index]);
 974                         spin_unlock(&dev->pending_lock);
 975                         /*
 976                          * If we get a partial ACK for a resent operation,
 977                          * we can stop resending the earlier packets and
 978                          * continue with the next packet the receiver wants.
 979                          */
 980                         if (ipath_cmp24(qp->s_psn, psn) <= 0) {
 981                                 reset_psn(qp, psn + 1);
 982                                 tasklet_hi_schedule(&qp->s_task);
 983                         }
 984                 } else if (ipath_cmp24(qp->s_psn, psn) <= 0) {
 985                         qp->s_state = OP(SEND_LAST);
 986                         qp->s_psn = psn + 1;
 987                 }
 988                 ipath_get_credit(qp, aeth);
 989                 qp->s_rnr_retry = qp->s_rnr_retry_cnt;
 990                 qp->s_retry = qp->s_retry_cnt;
 991                 update_last_psn(qp, psn);
 992                 ret = 1;
 993                 goto bail;
 994
 995         case 1:         /* RNR NAK */
 996                 dev->n_rnr_naks++;
 997                 if (qp->s_last == qp->s_tail)
 998                         goto bail;
 999                 if (qp->s_rnr_retry == 0) {
1000                         status = IB_WC_RNR_RETRY_EXC_ERR;
1001                         goto class_b;
1002                 }
1003                 if (qp->s_rnr_retry_cnt < 7)
1004                         qp->s_rnr_retry--;
1005
1006                 /* The last valid PSN is the previous PSN. */
1007                 update_last_psn(qp, psn - 1);
1008
1009                 if (wqe->wr.opcode == IB_WR_RDMA_READ)
1010                         dev->n_rc_resends++;
1011                 else
1012                         dev->n_rc_resends +=
1013                                 (qp->s_psn - psn) & IPATH_PSN_MASK;
1014
1015                 reset_psn(qp, psn);
1016
1017                 qp->s_rnr_timeout =
1018                         ib_ipath_rnr_table[(aeth >> IPATH_AETH_CREDIT_SHIFT) &
1019                                            IPATH_AETH_CREDIT_MASK];
1020                 ipath_insert_rnr_queue(qp);
1021                 goto bail;
1022
1023         case 3:         /* NAK */
1024                 if (qp->s_last == qp->s_tail)
1025                         goto bail;
1026                 /* The last valid PSN is the previous PSN. */
1027                 update_last_psn(qp, psn - 1);
1028                 switch ((aeth >> IPATH_AETH_CREDIT_SHIFT) &
1029                         IPATH_AETH_CREDIT_MASK) {
1030                 case 0: /* PSN sequence error */
1031                         dev->n_seq_naks++;
1032                         /*
1033                          * Back up to the responder's expected PSN.
1034                          * Note that we might get a NAK in the middle of an
1035                          * RDMA READ response which terminates the RDMA
1036                          * READ.
1037                          */
1038                         ipath_restart_rc(qp, psn);
1039                         break;
1040
1041                 case 1: /* Invalid Request */
1042                         status = IB_WC_REM_INV_REQ_ERR;
1043                         dev->n_other_naks++;
1044                         goto class_b;
1045
1046                 case 2: /* Remote Access Error */
1047                         status = IB_WC_REM_ACCESS_ERR;
1048                         dev->n_other_naks++;
1049                         goto class_b;
1050
1051                 case 3: /* Remote Operation Error */
1052                         status = IB_WC_REM_OP_ERR;
1053                         dev->n_other_naks++;
1054                 class_b:
1055                         ipath_send_complete(qp, wqe, status);
1056                         ipath_error_qp(qp, IB_WC_WR_FLUSH_ERR);
1057                         break;
1058
1059                 default:
1060                         /* Ignore other reserved NAK error codes */
1061                         goto reserved;
1062                 }
1063                 qp->s_rnr_retry = qp->s_rnr_retry_cnt;
1064                 goto bail;
1065
1066         default:                /* 2: reserved */
1067         reserved:
1068                 /* Ignore reserved NAK codes. */
1069                 goto bail;
1070         }
1071
1072 bail:
1073         return ret;
1074 }
1075
1076 /**
1077  * ipath_rc_rcv_resp - process an incoming RC response packet
1078  * @dev: the device this packet came in on
1079  * @ohdr: the other headers for this packet
1080  * @data: the packet data
1081  * @tlen: the packet length
1082  * @qp: the QP for this packet
1083  * @opcode: the opcode for this packet
1084  * @psn: the packet sequence number for this packet
1085  * @hdrsize: the header length
1086  * @pmtu: the path MTU
1087  * @header_in_data: true if part of the header data is in the data buffer
1088  *
1089  * This is called from ipath_rc_rcv() to process an incoming RC response
1090  * packet for the given QP.
1091  * Called at interrupt level.
1092  */
1093 static inline void ipath_rc_rcv_resp(struct ipath_ibdev *dev,
1094                                      struct ipath_other_headers *ohdr,
1095                                      void *data, u32 tlen,
1096                                      struct ipath_qp *qp,
1097                                      u32 opcode,
1098                                      u32 psn, u32 hdrsize, u32 pmtu,
1099                                      int header_in_data)
1100 {
1101         struct ipath_swqe *wqe;
1102         enum ib_wc_status status;
1103         unsigned long flags;
1104         int diff;
1105         u32 pad;
1106         u32 aeth;
1107         u64 val;
1108
1109         spin_lock_irqsave(&qp->s_lock, flags);
1110
1111         /* Ignore invalid responses. */
1112         if (ipath_cmp24(psn, qp->s_next_psn) >= 0)
1113                 goto ack_done;
1114
1115         /* Ignore duplicate responses. */
1116         diff = ipath_cmp24(psn, qp->s_last_psn);
1117         if (unlikely(diff <= 0)) {
1118                 /* Update credits for "ghost" ACKs */
1119                 if (diff == 0 && opcode == OP(ACKNOWLEDGE)) {
1120                         if (!header_in_data)
1121                                 aeth = be32_to_cpu(ohdr->u.aeth);
1122                         else {
1123                                 aeth = be32_to_cpu(((__be32 *) data)[0]);
1124                                 data += sizeof(__be32);
1125                         }
1126                         if ((aeth >> 29) == 0)
1127                                 ipath_get_credit(qp, aeth);
1128                 }
1129                 goto ack_done;
1130         }
1131
1132         if (unlikely(qp->s_last == qp->s_tail))
1133                 goto ack_done;
1134         wqe = get_swqe_ptr(qp, qp->s_last);
1135         status = IB_WC_SUCCESS;
1136
1137         switch (opcode) {
1138         case OP(ACKNOWLEDGE):
1139         case OP(ATOMIC_ACKNOWLEDGE):
1140         case OP(RDMA_READ_RESPONSE_FIRST):
1141                 if (!header_in_data)
1142                         aeth = be32_to_cpu(ohdr->u.aeth);
1143                 else {
1144                         aeth = be32_to_cpu(((__be32 *) data)[0]);
1145                         data += sizeof(__be32);
1146                 }
1147                 if (opcode == OP(ATOMIC_ACKNOWLEDGE)) {
1148                         if (!header_in_data) {
1149                                 __be32 *p = ohdr->u.at.atomic_ack_eth;
1150
1151                                 val = ((u64) be32_to_cpu(p[0]) << 32) |
1152                                         be32_to_cpu(p[1]);
1153                         } else
1154                                 val = be64_to_cpu(((__be64 *) data)[0]);
1155                 } else
1156                         val = 0;
1157                 if (!do_rc_ack(qp, aeth, psn, opcode, val) ||
1158                     opcode != OP(RDMA_READ_RESPONSE_FIRST))
1159                         goto ack_done;
1160                 hdrsize += 4;
1161                 wqe = get_swqe_ptr(qp, qp->s_last);
1162                 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1163                         goto ack_op_err;
1164                 /*
1165                  * If this is a response to a resent RDMA read, we
1166                  * have to be careful to copy the data to the right
1167                  * location.
1168                  */
1169                 qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
1170                                                   wqe, psn, pmtu);
1171                 goto read_middle;
1172
1173         case OP(RDMA_READ_RESPONSE_MIDDLE):
1174                 /* no AETH, no ACK */
1175                 if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) {
1176                         dev->n_rdma_seq++;
1177                         ipath_restart_rc(qp, qp->s_last_psn + 1);
1178                         goto ack_done;
1179                 }
1180                 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1181                         goto ack_op_err;
1182         read_middle:
1183                 if (unlikely(tlen != (hdrsize + pmtu + 4)))
1184                         goto ack_len_err;
1185                 if (unlikely(pmtu >= qp->s_rdma_read_len))
1186                         goto ack_len_err;
1187
1188                 /* We got a response so update the timeout. */
1189                 spin_lock(&dev->pending_lock);
1190                 if (qp->s_rnr_timeout == 0 && !list_empty(&qp->timerwait))
1191                         list_move_tail(&qp->timerwait,
1192                                        &dev->pending[dev->pending_index]);
1193                 spin_unlock(&dev->pending_lock);
1194
1195                 if (opcode == OP(RDMA_READ_RESPONSE_MIDDLE))
1196                         qp->s_retry = qp->s_retry_cnt;
1197
1198                 /*
1199                  * Update the RDMA receive state but do the copy w/o
1200                  * holding the locks and blocking interrupts.
1201                  */
1202                 qp->s_rdma_read_len -= pmtu;
1203                 update_last_psn(qp, psn);
1204                 spin_unlock_irqrestore(&qp->s_lock, flags);
1205                 ipath_copy_sge(&qp->s_rdma_read_sge, data, pmtu);
1206                 goto bail;
1207
1208         case OP(RDMA_READ_RESPONSE_ONLY):
1209                 if (!header_in_data)
1210                         aeth = be32_to_cpu(ohdr->u.aeth);
1211                 else
1212                         aeth = be32_to_cpu(((__be32 *) data)[0]);
1213                 if (!do_rc_ack(qp, aeth, psn, opcode, 0))
1214                         goto ack_done;
1215                 /* Get the number of bytes the message was padded by. */
1216                 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1217                 /*
1218                  * Check that the data size is >= 0 && <= pmtu.
1219                  * Remember to account for the AETH header (4) and
1220                  * ICRC (4).
1221                  */
1222                 if (unlikely(tlen < (hdrsize + pad + 8)))
1223                         goto ack_len_err;
1224                 /*
1225                  * If this is a response to a resent RDMA read, we
1226                  * have to be careful to copy the data to the right
1227                  * location.
1228                  */
1229                 wqe = get_swqe_ptr(qp, qp->s_last);
1230                 qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
1231                                                   wqe, psn, pmtu);
1232                 goto read_last;
1233
1234         case OP(RDMA_READ_RESPONSE_LAST):
1235                 /* ACKs READ req. */
1236                 if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) {
1237                         dev->n_rdma_seq++;
1238                         ipath_restart_rc(qp, qp->s_last_psn + 1);
1239                         goto ack_done;
1240                 }
1241                 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1242                         goto ack_op_err;
1243                 /* Get the number of bytes the message was padded by. */
1244                 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1245                 /*
1246                  * Check that the data size is >= 1 && <= pmtu.
1247                  * Remember to account for the AETH header (4) and
1248                  * ICRC (4).
1249                  */
1250                 if (unlikely(tlen <= (hdrsize + pad + 8)))
1251                         goto ack_len_err;
1252         read_last:
1253                 tlen -= hdrsize + pad + 8;
1254                 if (unlikely(tlen != qp->s_rdma_read_len))
1255                         goto ack_len_err;
1256                 if (!header_in_data)
1257                         aeth = be32_to_cpu(ohdr->u.aeth);
1258                 else {
1259                         aeth = be32_to_cpu(((__be32 *) data)[0]);
1260                         data += sizeof(__be32);
1261                 }
1262                 ipath_copy_sge(&qp->s_rdma_read_sge, data, tlen);
1263                 (void) do_rc_ack(qp, aeth, psn,
1264                                  OP(RDMA_READ_RESPONSE_LAST), 0);
1265                 goto ack_done;
1266         }
1267
1268 ack_op_err:
1269         status = IB_WC_LOC_QP_OP_ERR;
1270         goto ack_err;
1271
1272 ack_len_err:
1273         status = IB_WC_LOC_LEN_ERR;
1274 ack_err:
1275         ipath_send_complete(qp, wqe, status);
1276         ipath_error_qp(qp, IB_WC_WR_FLUSH_ERR);
1277 ack_done:
1278         spin_unlock_irqrestore(&qp->s_lock, flags);
1279 bail:
1280         return;
1281 }
1282
1283 /**
1284  * ipath_rc_rcv_error - process an incoming duplicate or error RC packet
1285  * @dev: the device this packet came in on
1286  * @ohdr: the other headers for this packet
1287  * @data: the packet data
1288  * @qp: the QP for this packet
1289  * @opcode: the opcode for this packet
1290  * @psn: the packet sequence number for this packet
1291  * @diff: the difference between the PSN and the expected PSN
1292  * @header_in_data: true if part of the header data is in the data buffer
1293  *
1294  * This is called from ipath_rc_rcv() to process an unexpected
1295  * incoming RC packet for the given QP.
1296  * Called at interrupt level.
1297  * Return 1 if no more processing is needed; otherwise return 0 to
1298  * schedule a response to be sent.
1299  */
1300 static inline int ipath_rc_rcv_error(struct ipath_ibdev *dev,
1301                                      struct ipath_other_headers *ohdr,
1302                                      void *data,
1303                                      struct ipath_qp *qp,
1304                                      u32 opcode,
1305                                      u32 psn,
1306                                      int diff,
1307                                      int header_in_data)
1308 {
1309         struct ipath_ack_entry *e;
1310         u8 i, prev;
1311         int old_req;
1312         unsigned long flags;
1313
1314         if (diff > 0) {
1315                 /*
1316                  * Packet sequence error.
1317                  * A NAK will ACK earlier sends and RDMA writes.
1318                  * Don't queue the NAK if we already sent one.
1319                  */
1320                 if (!qp->r_nak_state) {
1321                         qp->r_nak_state = IB_NAK_PSN_ERROR;
1322                         /* Use the expected PSN. */
1323                         qp->r_ack_psn = qp->r_psn;
1324                         goto send_ack;
1325                 }
1326                 goto done;
1327         }
1328
1329         /*
1330          * Handle a duplicate request.  Don't re-execute SEND, RDMA
1331          * write or atomic op.  Don't NAK errors, just silently drop
1332          * the duplicate request.  Note that r_sge, r_len, and
1333          * r_rcv_len may be in use so don't modify them.
1334          *
1335          * We are supposed to ACK the earliest duplicate PSN but we
1336          * can coalesce an outstanding duplicate ACK.  We have to
1337          * send the earliest so that RDMA reads can be restarted at
1338          * the requester's expected PSN.
1339          *
1340          * First, find where this duplicate PSN falls within the
1341          * ACKs previously sent.
1342          */
1343         psn &= IPATH_PSN_MASK;
1344         e = NULL;
1345         old_req = 1;
1346         spin_lock_irqsave(&qp->s_lock, flags);
1347         for (i = qp->r_head_ack_queue; ; i = prev) {
1348                 if (i == qp->s_tail_ack_queue)
1349                         old_req = 0;
1350                 if (i)
1351                         prev = i - 1;
1352                 else
1353                         prev = IPATH_MAX_RDMA_ATOMIC;
1354                 if (prev == qp->r_head_ack_queue) {
1355                         e = NULL;
1356                         break;
1357                 }
1358                 e = &qp->s_ack_queue[prev];
1359                 if (!e->opcode) {
1360                         e = NULL;
1361                         break;
1362                 }
1363                 if (ipath_cmp24(psn, e->psn) >= 0) {
1364                         if (prev == qp->s_tail_ack_queue)
1365                                 old_req = 0;
1366                         break;
1367                 }
1368         }
1369         switch (opcode) {
1370         case OP(RDMA_READ_REQUEST): {
1371                 struct ib_reth *reth;
1372                 u32 offset;
1373                 u32 len;
1374
1375                 /*
1376                  * If we didn't find the RDMA read request in the ack queue,
1377                  * or the send tasklet is already backed up to send an
1378                  * earlier entry, we can ignore this request.
1379                  */
1380                 if (!e || e->opcode != OP(RDMA_READ_REQUEST) || old_req)
1381                         goto unlock_done;
1382                 /* RETH comes after BTH */
1383                 if (!header_in_data)
1384                         reth = &ohdr->u.rc.reth;
1385                 else {
1386                         reth = (struct ib_reth *)data;
1387                         data += sizeof(*reth);
1388                 }
1389                 /*
1390                  * Address range must be a subset of the original
1391                  * request and start on pmtu boundaries.
1392                  * We reuse the old ack_queue slot since the requester
1393                  * should not back up and request an earlier PSN for the
1394                  * same request.
1395                  */
1396                 offset = ((psn - e->psn) & IPATH_PSN_MASK) *
1397                         ib_mtu_enum_to_int(qp->path_mtu);
1398                 len = be32_to_cpu(reth->length);
1399                 if (unlikely(offset + len > e->rdma_sge.sge.sge_length))
1400                         goto unlock_done;
1401                 if (len != 0) {
1402                         u32 rkey = be32_to_cpu(reth->rkey);
1403                         u64 vaddr = be64_to_cpu(reth->vaddr);
1404                         int ok;
1405
1406                         ok = ipath_rkey_ok(qp, &e->rdma_sge,
1407                                            len, vaddr, rkey,
1408                                            IB_ACCESS_REMOTE_READ);
1409                         if (unlikely(!ok))
1410                                 goto unlock_done;
1411                 } else {
1412                         e->rdma_sge.sg_list = NULL;
1413                         e->rdma_sge.num_sge = 0;
1414                         e->rdma_sge.sge.mr = NULL;
1415                         e->rdma_sge.sge.vaddr = NULL;
1416                         e->rdma_sge.sge.length = 0;
1417                         e->rdma_sge.sge.sge_length = 0;
1418                 }
1419                 e->psn = psn;
1420                 qp->s_ack_state = OP(ACKNOWLEDGE);
1421                 qp->s_tail_ack_queue = prev;
1422                 break;
1423         }
1424
1425         case OP(COMPARE_SWAP):
1426         case OP(FETCH_ADD): {
1427                 /*
1428                  * If we didn't find the atomic request in the ack queue
1429                  * or the send tasklet is already backed up to send an
1430                  * earlier entry, we can ignore this request.
1431                  */
1432                 if (!e || e->opcode != (u8) opcode || old_req)
1433                         goto unlock_done;
1434                 qp->s_ack_state = OP(ACKNOWLEDGE);
1435                 qp->s_tail_ack_queue = prev;
1436                 break;
1437         }
1438
1439         default:
1440                 if (old_req)
1441                         goto unlock_done;
1442                 /*
1443                  * Resend the most recent ACK if this request is
1444                  * after all the previous RDMA reads and atomics.
1445                  */
1446                 if (i == qp->r_head_ack_queue) {
1447                         spin_unlock_irqrestore(&qp->s_lock, flags);
1448                         qp->r_nak_state = 0;
1449                         qp->r_ack_psn = qp->r_psn - 1;
1450                         goto send_ack;
1451                 }
1452                 /*
1453                  * Try to send a simple ACK to work around a Mellanox bug
1454                  * which doesn't accept a RDMA read response or atomic
1455                  * response as an ACK for earlier SENDs or RDMA writes.
1456                  */
1457                 if (qp->r_head_ack_queue == qp->s_tail_ack_queue &&
1458                     !(qp->s_flags & IPATH_S_ACK_PENDING) &&
1459                     qp->s_ack_state == OP(ACKNOWLEDGE)) {
1460                         spin_unlock_irqrestore(&qp->s_lock, flags);
1461                         qp->r_nak_state = 0;
1462                         qp->r_ack_psn = qp->s_ack_queue[i].psn - 1;
1463                         goto send_ack;
1464                 }
1465                 /*
1466                  * Resend the RDMA read or atomic op which
1467                  * ACKs this duplicate request.
1468                  */
1469                 qp->s_ack_state = OP(ACKNOWLEDGE);
1470                 qp->s_tail_ack_queue = i;
1471                 break;
1472         }
1473         qp->r_nak_state = 0;
1474         tasklet_hi_schedule(&qp->s_task);
1475
1476 unlock_done:
1477         spin_unlock_irqrestore(&qp->s_lock, flags);
1478 done:
1479         return 1;
1480
1481 send_ack:
1482         return 0;
1483 }
1484
1485 void ipath_rc_error(struct ipath_qp *qp, enum ib_wc_status err)
1486 {
1487         unsigned long flags;
1488         int lastwqe;
1489
1490         spin_lock_irqsave(&qp->s_lock, flags);
1491         lastwqe = ipath_error_qp(qp, err);
1492         spin_unlock_irqrestore(&qp->s_lock, flags);
1493
1494         if (lastwqe) {
1495                 struct ib_event ev;
1496
1497                 ev.device = qp->ibqp.device;
1498                 ev.element.qp = &qp->ibqp;
1499                 ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
1500                 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
1501         }
1502 }
1503
1504 static inline void ipath_update_ack_queue(struct ipath_qp *qp, unsigned n)
1505 {
1506         unsigned long flags;
1507         unsigned next;
1508
1509         next = n + 1;
1510         if (next > IPATH_MAX_RDMA_ATOMIC)
1511                 next = 0;
1512         spin_lock_irqsave(&qp->s_lock, flags);
1513         if (n == qp->s_tail_ack_queue) {
1514                 qp->s_tail_ack_queue = next;
1515                 qp->s_ack_state = OP(ACKNOWLEDGE);
1516         }
1517         spin_unlock_irqrestore(&qp->s_lock, flags);
1518 }
1519
1520 /**
1521  * ipath_rc_rcv - process an incoming RC packet
1522  * @dev: the device this packet came in on
1523  * @hdr: the header of this packet
1524  * @has_grh: true if the header has a GRH
1525  * @data: the packet data
1526  * @tlen: the packet length
1527  * @qp: the QP for this packet
1528  *
1529  * This is called from ipath_qp_rcv() to process an incoming RC packet
1530  * for the given QP.
1531  * Called at interrupt level.
1532  */
1533 void ipath_rc_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
1534                   int has_grh, void *data, u32 tlen, struct ipath_qp *qp)
1535 {
1536         struct ipath_other_headers *ohdr;
1537         u32 opcode;
1538         u32 hdrsize;
1539         u32 psn;
1540         u32 pad;
1541         struct ib_wc wc;
1542         u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
1543         int diff;
1544         struct ib_reth *reth;
1545         int header_in_data;
1546
1547         /* Validate the SLID. See Ch. 9.6.1.5 */
1548         if (unlikely(be16_to_cpu(hdr->lrh[3]) != qp->remote_ah_attr.dlid))
1549                 goto done;
1550
1551         /* Check for GRH */
1552         if (!has_grh) {
1553                 ohdr = &hdr->u.oth;
1554                 hdrsize = 8 + 12;       /* LRH + BTH */
1555                 psn = be32_to_cpu(ohdr->bth[2]);
1556                 header_in_data = 0;
1557         } else {
1558                 ohdr = &hdr->u.l.oth;
1559                 hdrsize = 8 + 40 + 12;  /* LRH + GRH + BTH */
1560                 /*
1561                  * The header with GRH is 60 bytes and the core driver sets
1562                  * the eager header buffer size to 56 bytes so the last 4
1563                  * bytes of the BTH header (PSN) is in the data buffer.
1564                  */
1565                 header_in_data = dev->dd->ipath_rcvhdrentsize == 16;
1566                 if (header_in_data) {
1567                         psn = be32_to_cpu(((__be32 *) data)[0]);
1568                         data += sizeof(__be32);
1569                 } else
1570                         psn = be32_to_cpu(ohdr->bth[2]);
1571         }
1572
1573         /*
1574          * Process responses (ACKs) before anything else.  Note that the
1575          * packet sequence number will be for something in the send work
1576          * queue rather than the expected receive packet sequence number.
1577          * In other words, this QP is the requester.
1578          */
1579         opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
1580         if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
1581             opcode <= OP(ATOMIC_ACKNOWLEDGE)) {
1582                 ipath_rc_rcv_resp(dev, ohdr, data, tlen, qp, opcode, psn,
1583                                   hdrsize, pmtu, header_in_data);
1584                 goto done;
1585         }
1586
1587         /* Compute 24 bits worth of difference. */
1588         diff = ipath_cmp24(psn, qp->r_psn);
1589         if (unlikely(diff)) {
1590                 if (ipath_rc_rcv_error(dev, ohdr, data, qp, opcode,
1591                                        psn, diff, header_in_data))
1592                         goto done;
1593                 goto send_ack;
1594         }
1595
1596         /* Check for opcode sequence errors. */
1597         switch (qp->r_state) {
1598         case OP(SEND_FIRST):
1599         case OP(SEND_MIDDLE):
1600                 if (opcode == OP(SEND_MIDDLE) ||
1601                     opcode == OP(SEND_LAST) ||
1602                     opcode == OP(SEND_LAST_WITH_IMMEDIATE))
1603                         break;
1604                 goto nack_inv;
1605
1606         case OP(RDMA_WRITE_FIRST):
1607         case OP(RDMA_WRITE_MIDDLE):
1608                 if (opcode == OP(RDMA_WRITE_MIDDLE) ||
1609                     opcode == OP(RDMA_WRITE_LAST) ||
1610                     opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
1611                         break;
1612                 goto nack_inv;
1613
1614         default:
1615                 if (opcode == OP(SEND_MIDDLE) ||
1616                     opcode == OP(SEND_LAST) ||
1617                     opcode == OP(SEND_LAST_WITH_IMMEDIATE) ||
1618                     opcode == OP(RDMA_WRITE_MIDDLE) ||
1619                     opcode == OP(RDMA_WRITE_LAST) ||
1620                     opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
1621                         goto nack_inv;
1622                 /*
1623                  * Note that it is up to the requester to not send a new
1624                  * RDMA read or atomic operation before receiving an ACK
1625                  * for the previous operation.
1626                  */
1627                 break;
1628         }
1629
1630         memset(&wc, 0, sizeof wc);
1631
1632         /* OK, process the packet. */
1633         switch (opcode) {
1634         case OP(SEND_FIRST):
1635                 if (!ipath_get_rwqe(qp, 0))
1636                         goto rnr_nak;
1637                 qp->r_rcv_len = 0;
1638                 /* FALLTHROUGH */
1639         case OP(SEND_MIDDLE):
1640         case OP(RDMA_WRITE_MIDDLE):
1641         send_middle:
1642                 /* Check for invalid length PMTU or posted rwqe len. */
1643                 if (unlikely(tlen != (hdrsize + pmtu + 4)))
1644                         goto nack_inv;
1645                 qp->r_rcv_len += pmtu;
1646                 if (unlikely(qp->r_rcv_len > qp->r_len))
1647                         goto nack_inv;
1648                 ipath_copy_sge(&qp->r_sge, data, pmtu);
1649                 break;
1650
1651         case OP(RDMA_WRITE_LAST_WITH_IMMEDIATE):
1652                 /* consume RWQE */
1653                 if (!ipath_get_rwqe(qp, 1))
1654                         goto rnr_nak;
1655                 goto send_last_imm;
1656
1657         case OP(SEND_ONLY):
1658         case OP(SEND_ONLY_WITH_IMMEDIATE):
1659                 if (!ipath_get_rwqe(qp, 0))
1660                         goto rnr_nak;
1661                 qp->r_rcv_len = 0;
1662                 if (opcode == OP(SEND_ONLY))
1663                         goto send_last;
1664                 /* FALLTHROUGH */
1665         case OP(SEND_LAST_WITH_IMMEDIATE):
1666         send_last_imm:
1667                 if (header_in_data) {
1668                         wc.imm_data = *(__be32 *) data;
1669                         data += sizeof(__be32);
1670                 } else {
1671                         /* Immediate data comes after BTH */
1672                         wc.imm_data = ohdr->u.imm_data;
1673                 }
1674                 hdrsize += 4;
1675                 wc.wc_flags = IB_WC_WITH_IMM;
1676                 /* FALLTHROUGH */
1677         case OP(SEND_LAST):
1678         case OP(RDMA_WRITE_LAST):
1679         send_last:
1680                 /* Get the number of bytes the message was padded by. */
1681                 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1682                 /* Check for invalid length. */
1683                 /* XXX LAST len should be >= 1 */
1684                 if (unlikely(tlen < (hdrsize + pad + 4)))
1685                         goto nack_inv;
1686                 /* Don't count the CRC. */
1687                 tlen -= (hdrsize + pad + 4);
1688                 wc.byte_len = tlen + qp->r_rcv_len;
1689                 if (unlikely(wc.byte_len > qp->r_len))
1690                         goto nack_inv;
1691                 ipath_copy_sge(&qp->r_sge, data, tlen);
1692                 qp->r_msn++;
1693                 if (!qp->r_wrid_valid)
1694                         break;
1695                 qp->r_wrid_valid = 0;
1696                 wc.wr_id = qp->r_wr_id;
1697                 wc.status = IB_WC_SUCCESS;
1698                 if (opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE) ||
1699                     opcode == OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE))
1700                         wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
1701                 else
1702                         wc.opcode = IB_WC_RECV;
1703                 wc.qp = &qp->ibqp;
1704                 wc.src_qp = qp->remote_qpn;
1705                 wc.slid = qp->remote_ah_attr.dlid;
1706                 wc.sl = qp->remote_ah_attr.sl;
1707                 /* Signal completion event if the solicited bit is set. */
1708                 ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
1709                                (ohdr->bth[0] &
1710                                 __constant_cpu_to_be32(1 << 23)) != 0);
1711                 break;
1712
1713         case OP(RDMA_WRITE_FIRST):
1714         case OP(RDMA_WRITE_ONLY):
1715         case OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE):
1716                 if (unlikely(!(qp->qp_access_flags &
1717                                IB_ACCESS_REMOTE_WRITE)))
1718                         goto nack_inv;
1719                 /* consume RWQE */
1720                 /* RETH comes after BTH */
1721                 if (!header_in_data)
1722                         reth = &ohdr->u.rc.reth;
1723                 else {
1724                         reth = (struct ib_reth *)data;
1725                         data += sizeof(*reth);
1726                 }
1727                 hdrsize += sizeof(*reth);
1728                 qp->r_len = be32_to_cpu(reth->length);
1729                 qp->r_rcv_len = 0;
1730                 if (qp->r_len != 0) {
1731                         u32 rkey = be32_to_cpu(reth->rkey);
1732                         u64 vaddr = be64_to_cpu(reth->vaddr);
1733                         int ok;
1734
1735                         /* Check rkey & NAK */
1736                         ok = ipath_rkey_ok(qp, &qp->r_sge,
1737                                            qp->r_len, vaddr, rkey,
1738                                            IB_ACCESS_REMOTE_WRITE);
1739                         if (unlikely(!ok))
1740                                 goto nack_acc;
1741                 } else {
1742                         qp->r_sge.sg_list = NULL;
1743                         qp->r_sge.sge.mr = NULL;
1744                         qp->r_sge.sge.vaddr = NULL;
1745                         qp->r_sge.sge.length = 0;
1746                         qp->r_sge.sge.sge_length = 0;
1747                 }
1748                 if (opcode == OP(RDMA_WRITE_FIRST))
1749                         goto send_middle;
1750                 else if (opcode == OP(RDMA_WRITE_ONLY))
1751                         goto send_last;
1752                 if (!ipath_get_rwqe(qp, 1))
1753                         goto rnr_nak;
1754                 goto send_last_imm;
1755
1756         case OP(RDMA_READ_REQUEST): {
1757                 struct ipath_ack_entry *e;
1758                 u32 len;
1759                 u8 next;
1760
1761                 if (unlikely(!(qp->qp_access_flags &
1762                                IB_ACCESS_REMOTE_READ)))
1763                         goto nack_inv;
1764                 next = qp->r_head_ack_queue + 1;
1765                 if (next > IPATH_MAX_RDMA_ATOMIC)
1766                         next = 0;
1767                 if (unlikely(next == qp->s_tail_ack_queue)) {
1768                         if (!qp->s_ack_queue[next].sent)
1769                                 goto nack_inv;
1770                         ipath_update_ack_queue(qp, next);
1771                 }
1772                 e = &qp->s_ack_queue[qp->r_head_ack_queue];
1773                 /* RETH comes after BTH */
1774                 if (!header_in_data)
1775                         reth = &ohdr->u.rc.reth;
1776                 else {
1777                         reth = (struct ib_reth *)data;
1778                         data += sizeof(*reth);
1779                 }
1780                 len = be32_to_cpu(reth->length);
1781                 if (len) {
1782                         u32 rkey = be32_to_cpu(reth->rkey);
1783                         u64 vaddr = be64_to_cpu(reth->vaddr);
1784                         int ok;
1785
1786                         /* Check rkey & NAK */
1787                         ok = ipath_rkey_ok(qp, &e->rdma_sge, len, vaddr,
1788                                            rkey, IB_ACCESS_REMOTE_READ);
1789                         if (unlikely(!ok))
1790                                 goto nack_acc;
1791                         /*
1792                          * Update the next expected PSN.  We add 1 later
1793                          * below, so only add the remainder here.
1794                          */
1795                         if (len > pmtu)
1796                                 qp->r_psn += (len - 1) / pmtu;
1797                 } else {
1798                         e->rdma_sge.sg_list = NULL;
1799                         e->rdma_sge.num_sge = 0;
1800                         e->rdma_sge.sge.mr = NULL;
1801                         e->rdma_sge.sge.vaddr = NULL;
1802                         e->rdma_sge.sge.length = 0;
1803                         e->rdma_sge.sge.sge_length = 0;
1804                 }
1805                 e->opcode = opcode;
1806                 e->sent = 0;
1807                 e->psn = psn;
1808                 /*
1809                  * We need to increment the MSN here instead of when we
1810                  * finish sending the result since a duplicate request would
1811                  * increment it more than once.
1812                  */
1813                 qp->r_msn++;
1814                 qp->r_psn++;
1815                 qp->r_state = opcode;
1816                 qp->r_nak_state = 0;
1817                 barrier();
1818                 qp->r_head_ack_queue = next;
1819
1820                 /* Call ipath_do_rc_send() in another thread. */
1821                 tasklet_hi_schedule(&qp->s_task);
1822
1823                 goto done;
1824         }
1825
1826         case OP(COMPARE_SWAP):
1827         case OP(FETCH_ADD): {
1828                 struct ib_atomic_eth *ateth;
1829                 struct ipath_ack_entry *e;
1830                 u64 vaddr;
1831                 atomic64_t *maddr;
1832                 u64 sdata;
1833                 u32 rkey;
1834                 u8 next;
1835
1836                 if (unlikely(!(qp->qp_access_flags &
1837                                IB_ACCESS_REMOTE_ATOMIC)))
1838                         goto nack_inv;
1839                 next = qp->r_head_ack_queue + 1;
1840                 if (next > IPATH_MAX_RDMA_ATOMIC)
1841                         next = 0;
1842                 if (unlikely(next == qp->s_tail_ack_queue)) {
1843                         if (!qp->s_ack_queue[next].sent)
1844                                 goto nack_inv;
1845                         ipath_update_ack_queue(qp, next);
1846                 }
1847                 if (!header_in_data)
1848                         ateth = &ohdr->u.atomic_eth;
1849                 else
1850                         ateth = (struct ib_atomic_eth *)data;
1851                 vaddr = ((u64) be32_to_cpu(ateth->vaddr[0]) << 32) |
1852                         be32_to_cpu(ateth->vaddr[1]);
1853                 if (unlikely(vaddr & (sizeof(u64) - 1)))
1854                         goto nack_inv;
1855                 rkey = be32_to_cpu(ateth->rkey);
1856                 /* Check rkey & NAK */
1857                 if (unlikely(!ipath_rkey_ok(qp, &qp->r_sge,
1858                                             sizeof(u64), vaddr, rkey,
1859                                             IB_ACCESS_REMOTE_ATOMIC)))
1860                         goto nack_acc;
1861                 /* Perform atomic OP and save result. */
1862                 maddr = (atomic64_t *) qp->r_sge.sge.vaddr;
1863                 sdata = be64_to_cpu(ateth->swap_data);
1864                 e = &qp->s_ack_queue[qp->r_head_ack_queue];
1865                 e->atomic_data = (opcode == OP(FETCH_ADD)) ?
1866                         (u64) atomic64_add_return(sdata, maddr) - sdata :
1867                         (u64) cmpxchg((u64 *) qp->r_sge.sge.vaddr,
1868                                       be64_to_cpu(ateth->compare_data),
1869                                       sdata);
1870                 e->opcode = opcode;
1871                 e->sent = 0;
1872                 e->psn = psn & IPATH_PSN_MASK;
1873                 qp->r_msn++;
1874                 qp->r_psn++;
1875                 qp->r_state = opcode;
1876                 qp->r_nak_state = 0;
1877                 barrier();
1878                 qp->r_head_ack_queue = next;
1879
1880                 /* Call ipath_do_rc_send() in another thread. */
1881                 tasklet_hi_schedule(&qp->s_task);
1882
1883                 goto done;
1884         }
1885
1886         default:
1887                 /* NAK unknown opcodes. */
1888                 goto nack_inv;
1889         }
1890         qp->r_psn++;
1891         qp->r_state = opcode;
1892         qp->r_ack_psn = psn;
1893         qp->r_nak_state = 0;
1894         /* Send an ACK if requested or required. */
1895         if (psn & (1 << 31))
1896                 goto send_ack;
1897         goto done;
1898
1899 rnr_nak:
1900         qp->r_nak_state = IB_RNR_NAK | qp->r_min_rnr_timer;
1901         qp->r_ack_psn = qp->r_psn;
1902         goto send_ack;
1903
1904 nack_inv:
1905         ipath_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
1906         qp->r_nak_state = IB_NAK_INVALID_REQUEST;
1907         qp->r_ack_psn = qp->r_psn;
1908         goto send_ack;
1909
1910 nack_acc:
1911         ipath_rc_error(qp, IB_WC_LOC_PROT_ERR);
1912         qp->r_nak_state = IB_NAK_REMOTE_ACCESS_ERROR;
1913         qp->r_ack_psn = qp->r_psn;
1914 send_ack:
1915         send_rc_ack(qp);
1916
1917 done:
1918         return;
1919 }