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         /* Don't send an ACK if we aren't supposed to. */
  96         if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_RECV_OK))
  97                 goto bail;
  98
  99         /* header size in 32-bit words LRH+BTH = (8+12)/4. */
 100         hwords = 5;
 101
 102         switch (qp->s_ack_state) {
 103         case OP(RDMA_READ_RESPONSE_LAST):
 104         case OP(RDMA_READ_RESPONSE_ONLY):
 105         case OP(ATOMIC_ACKNOWLEDGE):
 106                 /*
 107                  * We can increment the tail pointer now that the last
 108                  * response has been sent instead of only being
 109                  * constructed.
 110                  */
 111                 if (++qp->s_tail_ack_queue > IPATH_MAX_RDMA_ATOMIC)
 112                         qp->s_tail_ack_queue = 0;
 113                 /* FALLTHROUGH */
 114         case OP(SEND_ONLY):
 115         case OP(ACKNOWLEDGE):
 116                 /* Check for no next entry in the queue. */
 117                 if (qp->r_head_ack_queue == qp->s_tail_ack_queue) {
 118                         if (qp->s_flags & IPATH_S_ACK_PENDING)
 119                                 goto normal;
 120                         qp->s_ack_state = OP(ACKNOWLEDGE);
 121                         goto bail;
 122                 }
 123
 124                 e = &qp->s_ack_queue[qp->s_tail_ack_queue];
 125                 if (e->opcode == OP(RDMA_READ_REQUEST)) {
 126                         /* Copy SGE state in case we need to resend */
 127                         qp->s_ack_rdma_sge = e->rdma_sge;
 128                         qp->s_cur_sge = &qp->s_ack_rdma_sge;
 129                         len = e->rdma_sge.sge.sge_length;
 130                         if (len > pmtu) {
 131                                 len = pmtu;
 132                                 qp->s_ack_state = OP(RDMA_READ_RESPONSE_FIRST);
 133                         } else {
 134                                 qp->s_ack_state = OP(RDMA_READ_RESPONSE_ONLY);
 135                                 e->sent = 1;
 136                         }
 137                         ohdr->u.aeth = ipath_compute_aeth(qp);
 138                         hwords++;
 139                         qp->s_ack_rdma_psn = e->psn;
 140                         bth2 = qp->s_ack_rdma_psn++ & IPATH_PSN_MASK;
 141                 } else {
 142                         /* COMPARE_SWAP or FETCH_ADD */
 143                         qp->s_cur_sge = NULL;
 144                         len = 0;
 145                         qp->s_ack_state = OP(ATOMIC_ACKNOWLEDGE);
 146                         ohdr->u.at.aeth = ipath_compute_aeth(qp);
 147                         ohdr->u.at.atomic_ack_eth[0] =
 148                                 cpu_to_be32(e->atomic_data >> 32);
 149                         ohdr->u.at.atomic_ack_eth[1] =
 150                                 cpu_to_be32(e->atomic_data);
 151                         hwords += sizeof(ohdr->u.at) / sizeof(u32);
 152                         bth2 = e->psn;
 153                         e->sent = 1;
 154                 }
 155                 bth0 = qp->s_ack_state << 24;
 156                 break;
 157
 158         case OP(RDMA_READ_RESPONSE_FIRST):
 159                 qp->s_ack_state = OP(RDMA_READ_RESPONSE_MIDDLE);
 160                 /* FALLTHROUGH */
 161         case OP(RDMA_READ_RESPONSE_MIDDLE):
 162                 len = qp->s_ack_rdma_sge.sge.sge_length;
 163                 if (len > pmtu)
 164                         len = pmtu;
 165                 else {
 166                         ohdr->u.aeth = ipath_compute_aeth(qp);
 167                         hwords++;
 168                         qp->s_ack_state = OP(RDMA_READ_RESPONSE_LAST);
 169                         qp->s_ack_queue[qp->s_tail_ack_queue].sent = 1;
 170                 }
 171                 bth0 = qp->s_ack_state << 24;
 172                 bth2 = qp->s_ack_rdma_psn++ & IPATH_PSN_MASK;
 173                 break;
 174
 175         default:
 176         normal:
 177                 /*
 178                  * Send a regular ACK.
 179                  * Set the s_ack_state so we wait until after sending
 180                  * the ACK before setting s_ack_state to ACKNOWLEDGE
 181                  * (see above).
 182                  */
 183                 qp->s_ack_state = OP(SEND_ONLY);
 184                 qp->s_flags &= ~IPATH_S_ACK_PENDING;
 185                 qp->s_cur_sge = NULL;
 186                 if (qp->s_nak_state)
 187                         ohdr->u.aeth =
 188                                 cpu_to_be32((qp->r_msn & IPATH_MSN_MASK) |
 189                                             (qp->s_nak_state <<
 190                                              IPATH_AETH_CREDIT_SHIFT));
 191                 else
 192                         ohdr->u.aeth = ipath_compute_aeth(qp);
 193                 hwords++;
 194                 len = 0;
 195                 bth0 = OP(ACKNOWLEDGE) << 24;
 196                 bth2 = qp->s_ack_psn & IPATH_PSN_MASK;
 197         }
 198         qp->s_hdrwords = hwords;
 199         qp->s_cur_size = len;
 200         ipath_make_ruc_header(dev, qp, ohdr, bth0, bth2);
 201         return 1;
 202
 203 bail:
 204         return 0;
 205 }
 206
 207 /**
 208  * ipath_make_rc_req - construct a request packet (SEND, RDMA r/w, ATOMIC)
 209  * @qp: a pointer to the QP
 210  *
 211  * Return 1 if constructed; otherwise, return 0.
 212  */
 213 int ipath_make_rc_req(struct ipath_qp *qp)
 214 {
 215         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
 216         struct ipath_other_headers *ohdr;
 217         struct ipath_sge_state *ss;
 218         struct ipath_swqe *wqe;
 219         u32 hwords;
 220         u32 len;
 221         u32 bth0;
 222         u32 bth2;
 223         u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
 224         char newreq;
 225         unsigned long flags;
 226         int ret = 0;
 227
 228         ohdr = &qp->s_hdr.u.oth;
 229         if (qp->remote_ah_attr.ah_flags & IB_AH_GRH)
 230                 ohdr = &qp->s_hdr.u.l.oth;
 231
 232         /*
 233          * The lock is needed to synchronize between the sending tasklet,
 234          * the receive interrupt handler, and timeout resends.
 235          */
 236         spin_lock_irqsave(&qp->s_lock, flags);
 237
 238         /* Sending responses has higher priority over sending requests. */
 239         if ((qp->r_head_ack_queue != qp->s_tail_ack_queue ||
 240              (qp->s_flags & IPATH_S_ACK_PENDING) ||
 241              qp->s_ack_state != OP(ACKNOWLEDGE)) &&
 242             ipath_make_rc_ack(dev, qp, ohdr, pmtu))
 243                 goto done;
 244
 245         if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK)) {
 246                 if (!(ib_ipath_state_ops[qp->state] & IPATH_FLUSH_SEND))
 247                         goto bail;
 248                 /* We are in the error state, flush the work request. */
 249                 if (qp->s_last == qp->s_head)
 250                         goto bail;
 251                 /* If DMAs are in progress, we can't flush immediately. */
 252                 if (atomic_read(&qp->s_dma_busy)) {
 253                         qp->s_flags |= IPATH_S_WAIT_DMA;
 254                         goto bail;
 255                 }
 256                 wqe = get_swqe_ptr(qp, qp->s_last);
 257                 ipath_send_complete(qp, wqe, IB_WC_WR_FLUSH_ERR);
 258                 goto done;
 259         }
 260
 261         /* Leave BUSY set until RNR timeout. */
 262         if (qp->s_rnr_timeout) {
 263                 qp->s_flags |= IPATH_S_WAITING;
 264                 goto bail;
 265         }
 266
 267         /* header size in 32-bit words LRH+BTH = (8+12)/4. */
 268         hwords = 5;
 269         bth0 = 1 << 22; /* Set M bit */
 270
 271         /* Send a request. */
 272         wqe = get_swqe_ptr(qp, qp->s_cur);
 273         switch (qp->s_state) {
 274         default:
 275                 if (!(ib_ipath_state_ops[qp->state] &
 276                     IPATH_PROCESS_NEXT_SEND_OK))
 277                         goto bail;
 278                 /*
 279                  * Resend an old request or start a new one.
 280                  *
 281                  * We keep track of the current SWQE so that
 282                  * we don't reset the "furthest progress" state
 283                  * if we need to back up.
 284                  */
 285                 newreq = 0;
 286                 if (qp->s_cur == qp->s_tail) {
 287                         /* Check if send work queue is empty. */
 288                         if (qp->s_tail == qp->s_head)
 289                                 goto bail;
 290                         /*
 291                          * If a fence is requested, wait for previous
 292                          * RDMA read and atomic operations to finish.
 293                          */
 294                         if ((wqe->wr.send_flags & IB_SEND_FENCE) &&
 295                             qp->s_num_rd_atomic) {
 296                                 qp->s_flags |= IPATH_S_FENCE_PENDING;
 297                                 goto bail;
 298                         }
 299                         wqe->psn = qp->s_next_psn;
 300                         newreq = 1;
 301                 }
 302                 /*
 303                  * Note that we have to be careful not to modify the
 304                  * original work request since we may need to resend
 305                  * it.
 306                  */
 307                 len = wqe->length;
 308                 ss = &qp->s_sge;
 309                 bth2 = 0;
 310                 switch (wqe->wr.opcode) {
 311                 case IB_WR_SEND:
 312                 case IB_WR_SEND_WITH_IMM:
 313                         /* If no credit, return. */
 314                         if (qp->s_lsn != (u32) -1 &&
 315                             ipath_cmp24(wqe->ssn, qp->s_lsn + 1) > 0) {
 316                                 qp->s_flags |= IPATH_S_WAIT_SSN_CREDIT;
 317                                 goto bail;
 318                         }
 319                         wqe->lpsn = wqe->psn;
 320                         if (len > pmtu) {
 321                                 wqe->lpsn += (len - 1) / pmtu;
 322                                 qp->s_state = OP(SEND_FIRST);
 323                                 len = pmtu;
 324                                 break;
 325                         }
 326                         if (wqe->wr.opcode == IB_WR_SEND)
 327                                 qp->s_state = OP(SEND_ONLY);
 328                         else {
 329                                 qp->s_state = OP(SEND_ONLY_WITH_IMMEDIATE);
 330                                 /* Immediate data comes after the BTH */
 331                                 ohdr->u.imm_data = wqe->wr.ex.imm_data;
 332                                 hwords += 1;
 333                         }
 334                         if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 335                                 bth0 |= 1 << 23;
 336                         bth2 = 1 << 31; /* Request ACK. */
 337                         if (++qp->s_cur == qp->s_size)
 338                                 qp->s_cur = 0;
 339                         break;
 340
 341                 case IB_WR_RDMA_WRITE:
 342                         if (newreq && qp->s_lsn != (u32) -1)
 343                                 qp->s_lsn++;
 344                         /* FALLTHROUGH */
 345                 case IB_WR_RDMA_WRITE_WITH_IMM:
 346                         /* If no credit, return. */
 347                         if (qp->s_lsn != (u32) -1 &&
 348                             ipath_cmp24(wqe->ssn, qp->s_lsn + 1) > 0) {
 349                                 qp->s_flags |= IPATH_S_WAIT_SSN_CREDIT;
 350                                 goto bail;
 351                         }
 352                         ohdr->u.rc.reth.vaddr =
 353                                 cpu_to_be64(wqe->wr.wr.rdma.remote_addr);
 354                         ohdr->u.rc.reth.rkey =
 355                                 cpu_to_be32(wqe->wr.wr.rdma.rkey);
 356                         ohdr->u.rc.reth.length = cpu_to_be32(len);
 357                         hwords += sizeof(struct ib_reth) / sizeof(u32);
 358                         wqe->lpsn = wqe->psn;
 359                         if (len > pmtu) {
 360                                 wqe->lpsn += (len - 1) / pmtu;
 361                                 qp->s_state = OP(RDMA_WRITE_FIRST);
 362                                 len = pmtu;
 363                                 break;
 364                         }
 365                         if (wqe->wr.opcode == IB_WR_RDMA_WRITE)
 366                                 qp->s_state = OP(RDMA_WRITE_ONLY);
 367                         else {
 368                                 qp->s_state =
 369                                         OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE);
 370                                 /* Immediate data comes after RETH */
 371                                 ohdr->u.rc.imm_data = wqe->wr.ex.imm_data;
 372                                 hwords += 1;
 373                                 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 374                                         bth0 |= 1 << 23;
 375                         }
 376                         bth2 = 1 << 31; /* Request ACK. */
 377                         if (++qp->s_cur == qp->s_size)
 378                                 qp->s_cur = 0;
 379                         break;
 380
 381                 case IB_WR_RDMA_READ:
 382                         /*
 383                          * Don't allow more operations to be started
 384                          * than the QP limits allow.
 385                          */
 386                         if (newreq) {
 387                                 if (qp->s_num_rd_atomic >=
 388                                     qp->s_max_rd_atomic) {
 389                                         qp->s_flags |= IPATH_S_RDMAR_PENDING;
 390                                         goto bail;
 391                                 }
 392                                 qp->s_num_rd_atomic++;
 393                                 if (qp->s_lsn != (u32) -1)
 394                                         qp->s_lsn++;
 395                                 /*
 396                                  * Adjust s_next_psn to count the
 397                                  * expected number of responses.
 398                                  */
 399                                 if (len > pmtu)
 400                                         qp->s_next_psn += (len - 1) / pmtu;
 401                                 wqe->lpsn = qp->s_next_psn++;
 402                         }
 403                         ohdr->u.rc.reth.vaddr =
 404                                 cpu_to_be64(wqe->wr.wr.rdma.remote_addr);
 405                         ohdr->u.rc.reth.rkey =
 406                                 cpu_to_be32(wqe->wr.wr.rdma.rkey);
 407                         ohdr->u.rc.reth.length = cpu_to_be32(len);
 408                         qp->s_state = OP(RDMA_READ_REQUEST);
 409                         hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
 410                         ss = NULL;
 411                         len = 0;
 412                         if (++qp->s_cur == qp->s_size)
 413                                 qp->s_cur = 0;
 414                         break;
 415
 416                 case IB_WR_ATOMIC_CMP_AND_SWP:
 417                 case IB_WR_ATOMIC_FETCH_AND_ADD:
 418                         /*
 419                          * Don't allow more operations to be started
 420                          * than the QP limits allow.
 421                          */
 422                         if (newreq) {
 423                                 if (qp->s_num_rd_atomic >=
 424                                     qp->s_max_rd_atomic) {
 425                                         qp->s_flags |= IPATH_S_RDMAR_PENDING;
 426                                         goto bail;
 427                                 }
 428                                 qp->s_num_rd_atomic++;
 429                                 if (qp->s_lsn != (u32) -1)
 430                                         qp->s_lsn++;
 431                                 wqe->lpsn = wqe->psn;
 432                         }
 433                         if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
 434                                 qp->s_state = OP(COMPARE_SWAP);
 435                                 ohdr->u.atomic_eth.swap_data = cpu_to_be64(
 436                                         wqe->wr.wr.atomic.swap);
 437                                 ohdr->u.atomic_eth.compare_data = cpu_to_be64(
 438                                         wqe->wr.wr.atomic.compare_add);
 439                         } else {
 440                                 qp->s_state = OP(FETCH_ADD);
 441                                 ohdr->u.atomic_eth.swap_data = cpu_to_be64(
 442                                         wqe->wr.wr.atomic.compare_add);
 443                                 ohdr->u.atomic_eth.compare_data = 0;
 444                         }
 445                         ohdr->u.atomic_eth.vaddr[0] = cpu_to_be32(
 446                                 wqe->wr.wr.atomic.remote_addr >> 32);
 447                         ohdr->u.atomic_eth.vaddr[1] = cpu_to_be32(
 448                                 wqe->wr.wr.atomic.remote_addr);
 449                         ohdr->u.atomic_eth.rkey = cpu_to_be32(
 450                                 wqe->wr.wr.atomic.rkey);
 451                         hwords += sizeof(struct ib_atomic_eth) / sizeof(u32);
 452                         ss = NULL;
 453                         len = 0;
 454                         if (++qp->s_cur == qp->s_size)
 455                                 qp->s_cur = 0;
 456                         break;
 457
 458                 default:
 459                         goto bail;
 460                 }
 461                 qp->s_sge.sge = wqe->sg_list[0];
 462                 qp->s_sge.sg_list = wqe->sg_list + 1;
 463                 qp->s_sge.num_sge = wqe->wr.num_sge;
 464                 qp->s_len = wqe->length;
 465                 if (newreq) {
 466                         qp->s_tail++;
 467                         if (qp->s_tail >= qp->s_size)
 468                                 qp->s_tail = 0;
 469                 }
 470                 bth2 |= qp->s_psn & IPATH_PSN_MASK;
 471                 if (wqe->wr.opcode == IB_WR_RDMA_READ)
 472                         qp->s_psn = wqe->lpsn + 1;
 473                 else {
 474                         qp->s_psn++;
 475                         if (ipath_cmp24(qp->s_psn, qp->s_next_psn) > 0)
 476                                 qp->s_next_psn = qp->s_psn;
 477                 }
 478                 /*
 479                  * Put the QP on the pending list so lost ACKs will cause
 480                  * a retry.  More than one request can be pending so the
 481                  * QP may already be on the dev->pending list.
 482                  */
 483                 spin_lock(&dev->pending_lock);
 484                 if (list_empty(&qp->timerwait))
 485                         list_add_tail(&qp->timerwait,
 486                                       &dev->pending[dev->pending_index]);
 487                 spin_unlock(&dev->pending_lock);
 488                 break;
 489
 490         case OP(RDMA_READ_RESPONSE_FIRST):
 491                 /*
 492                  * This case can only happen if a send is restarted.
 493                  * See ipath_restart_rc().
 494                  */
 495                 ipath_init_restart(qp, wqe);
 496                 /* FALLTHROUGH */
 497         case OP(SEND_FIRST):
 498                 qp->s_state = OP(SEND_MIDDLE);
 499                 /* FALLTHROUGH */
 500         case OP(SEND_MIDDLE):
 501                 bth2 = qp->s_psn++ & IPATH_PSN_MASK;
 502                 if (ipath_cmp24(qp->s_psn, qp->s_next_psn) > 0)
 503                         qp->s_next_psn = qp->s_psn;
 504                 ss = &qp->s_sge;
 505                 len = qp->s_len;
 506                 if (len > pmtu) {
 507                         len = pmtu;
 508                         break;
 509                 }
 510                 if (wqe->wr.opcode == IB_WR_SEND)
 511                         qp->s_state = OP(SEND_LAST);
 512                 else {
 513                         qp->s_state = OP(SEND_LAST_WITH_IMMEDIATE);
 514                         /* Immediate data comes after the BTH */
 515                         ohdr->u.imm_data = wqe->wr.ex.imm_data;
 516                         hwords += 1;
 517                 }
 518                 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 519                         bth0 |= 1 << 23;
 520                 bth2 |= 1 << 31;        /* Request ACK. */
 521                 qp->s_cur++;
 522                 if (qp->s_cur >= qp->s_size)
 523                         qp->s_cur = 0;
 524                 break;
 525
 526         case OP(RDMA_READ_RESPONSE_LAST):
 527                 /*
 528                  * This case can only happen if a RDMA write is restarted.
 529                  * See ipath_restart_rc().
 530                  */
 531                 ipath_init_restart(qp, wqe);
 532                 /* FALLTHROUGH */
 533         case OP(RDMA_WRITE_FIRST):
 534                 qp->s_state = OP(RDMA_WRITE_MIDDLE);
 535                 /* FALLTHROUGH */
 536         case OP(RDMA_WRITE_MIDDLE):
 537                 bth2 = qp->s_psn++ & IPATH_PSN_MASK;
 538                 if (ipath_cmp24(qp->s_psn, qp->s_next_psn) > 0)
 539                         qp->s_next_psn = qp->s_psn;
 540                 ss = &qp->s_sge;
 541                 len = qp->s_len;
 542                 if (len > pmtu) {
 543                         len = pmtu;
 544                         break;
 545                 }
 546                 if (wqe->wr.opcode == IB_WR_RDMA_WRITE)
 547                         qp->s_state = OP(RDMA_WRITE_LAST);
 548                 else {
 549                         qp->s_state = OP(RDMA_WRITE_LAST_WITH_IMMEDIATE);
 550                         /* Immediate data comes after the BTH */
 551                         ohdr->u.imm_data = wqe->wr.ex.imm_data;
 552                         hwords += 1;
 553                         if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 554                                 bth0 |= 1 << 23;
 555                 }
 556                 bth2 |= 1 << 31;        /* Request ACK. */
 557                 qp->s_cur++;
 558                 if (qp->s_cur >= qp->s_size)
 559                         qp->s_cur = 0;
 560                 break;
 561
 562         case OP(RDMA_READ_RESPONSE_MIDDLE):
 563                 /*
 564                  * This case can only happen if a RDMA read is restarted.
 565                  * See ipath_restart_rc().
 566                  */
 567                 ipath_init_restart(qp, wqe);
 568                 len = ((qp->s_psn - wqe->psn) & IPATH_PSN_MASK) * pmtu;
 569                 ohdr->u.rc.reth.vaddr =
 570                         cpu_to_be64(wqe->wr.wr.rdma.remote_addr + len);
 571                 ohdr->u.rc.reth.rkey =
 572                         cpu_to_be32(wqe->wr.wr.rdma.rkey);
 573                 ohdr->u.rc.reth.length = cpu_to_be32(qp->s_len);
 574                 qp->s_state = OP(RDMA_READ_REQUEST);
 575                 hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
 576                 bth2 = qp->s_psn++ & IPATH_PSN_MASK;
 577                 if (ipath_cmp24(qp->s_psn, qp->s_next_psn) > 0)
 578                         qp->s_next_psn = qp->s_psn;
 579                 ss = NULL;
 580                 len = 0;
 581                 qp->s_cur++;
 582                 if (qp->s_cur == qp->s_size)
 583                         qp->s_cur = 0;
 584                 break;
 585         }
 586         if (ipath_cmp24(qp->s_psn, qp->s_last_psn + IPATH_PSN_CREDIT - 1) >= 0)
 587                 bth2 |= 1 << 31;        /* Request ACK. */
 588         qp->s_len -= len;
 589         qp->s_hdrwords = hwords;
 590         qp->s_cur_sge = ss;
 591         qp->s_cur_size = len;
 592         ipath_make_ruc_header(dev, qp, ohdr, bth0 | (qp->s_state << 24), bth2);
 593 done:
 594         ret = 1;
 595         goto unlock;
 596
 597 bail:
 598         qp->s_flags &= ~IPATH_S_BUSY;
 599 unlock:
 600         spin_unlock_irqrestore(&qp->s_lock, flags);
 601         return ret;
 602 }
 603
 604 /**
 605  * send_rc_ack - Construct an ACK packet and send it
 606  * @qp: a pointer to the QP
 607  *
 608  * This is called from ipath_rc_rcv() and only uses the receive
 609  * side QP state.
 610  * Note that RDMA reads and atomics are handled in the
 611  * send side QP state and tasklet.
 612  */
 613 static void send_rc_ack(struct ipath_qp *qp)
 614 {
 615         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
 616         struct ipath_devdata *dd;
 617         u16 lrh0;
 618         u32 bth0;
 619         u32 hwords;
 620         u32 __iomem *piobuf;
 621         struct ipath_ib_header hdr;
 622         struct ipath_other_headers *ohdr;
 623         unsigned long flags;
 624
 625         spin_lock_irqsave(&qp->s_lock, flags);
 626
 627         /* Don't send ACK or NAK if a RDMA read or atomic is pending. */
 628         if (qp->r_head_ack_queue != qp->s_tail_ack_queue ||
 629             (qp->s_flags & IPATH_S_ACK_PENDING) ||
 630             qp->s_ack_state != OP(ACKNOWLEDGE))
 631                 goto queue_ack;
 632
 633         spin_unlock_irqrestore(&qp->s_lock, flags);
 634
 635         /* Don't try to send ACKs if the link isn't ACTIVE */
 636         dd = dev->dd;
 637         if (!(dd->ipath_flags & IPATH_LINKACTIVE))
 638                 goto done;
 639
 640         piobuf = ipath_getpiobuf(dd, 0, NULL);
 641         if (!piobuf) {
 642                 /*
 643                  * We are out of PIO buffers at the moment.
 644                  * Pass responsibility for sending the ACK to the
 645                  * send tasklet so that when a PIO buffer becomes
 646                  * available, the ACK is sent ahead of other outgoing
 647                  * packets.
 648                  */
 649                 spin_lock_irqsave(&qp->s_lock, flags);
 650                 goto queue_ack;
 651         }
 652
 653         /* Construct the header. */
 654         ohdr = &hdr.u.oth;
 655         lrh0 = IPATH_LRH_BTH;
 656         /* header size in 32-bit words LRH+BTH+AETH = (8+12+4)/4. */
 657         hwords = 6;
 658         if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
 659                 hwords += ipath_make_grh(dev, &hdr.u.l.grh,
 660                                          &qp->remote_ah_attr.grh,
 661                                          hwords, 0);
 662                 ohdr = &hdr.u.l.oth;
 663                 lrh0 = IPATH_LRH_GRH;
 664         }
 665         /* read pkey_index w/o lock (its atomic) */
 666         bth0 = ipath_get_pkey(dd, qp->s_pkey_index) |
 667                 (OP(ACKNOWLEDGE) << 24) | (1 << 22);
 668         if (qp->r_nak_state)
 669                 ohdr->u.aeth = cpu_to_be32((qp->r_msn & IPATH_MSN_MASK) |
 670                                             (qp->r_nak_state <<
 671                                              IPATH_AETH_CREDIT_SHIFT));
 672         else
 673                 ohdr->u.aeth = ipath_compute_aeth(qp);
 674         lrh0 |= qp->remote_ah_attr.sl << 4;
 675         hdr.lrh[0] = cpu_to_be16(lrh0);
 676         hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
 677         hdr.lrh[2] = cpu_to_be16(hwords + SIZE_OF_CRC);
 678         hdr.lrh[3] = cpu_to_be16(dd->ipath_lid |
 679                                  qp->remote_ah_attr.src_path_bits);
 680         ohdr->bth[0] = cpu_to_be32(bth0);
 681         ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
 682         ohdr->bth[2] = cpu_to_be32(qp->r_ack_psn & IPATH_PSN_MASK);
 683
 684         writeq(hwords + 1, piobuf);
 685
 686         if (dd->ipath_flags & IPATH_PIO_FLUSH_WC) {
 687                 u32 *hdrp = (u32 *) &hdr;
 688
 689                 ipath_flush_wc();
 690                 __iowrite32_copy(piobuf + 2, hdrp, hwords - 1);
 691                 ipath_flush_wc();
 692                 __raw_writel(hdrp[hwords - 1], piobuf + hwords + 1);
 693         } else
 694                 __iowrite32_copy(piobuf + 2, (u32 *) &hdr, hwords);
 695
 696         ipath_flush_wc();
 697
 698         dev->n_unicast_xmit++;
 699         goto done;
 700
 701 queue_ack:
 702         if (ib_ipath_state_ops[qp->state] & IPATH_PROCESS_RECV_OK) {
 703                 dev->n_rc_qacks++;
 704                 qp->s_flags |= IPATH_S_ACK_PENDING;
 705                 qp->s_nak_state = qp->r_nak_state;
 706                 qp->s_ack_psn = qp->r_ack_psn;
 707
 708                 /* Schedule the send tasklet. */
 709                 ipath_schedule_send(qp);
 710         }
 711         spin_unlock_irqrestore(&qp->s_lock, flags);
 712 done:
 713         return;
 714 }
 715
 716 /**
 717  * reset_psn - reset the QP state to send starting from PSN
 718  * @qp: the QP
 719  * @psn: the packet sequence number to restart at
 720  *
 721  * This is called from ipath_rc_rcv() to process an incoming RC ACK
 722  * for the given QP.
 723  * Called at interrupt level with the QP s_lock held.
 724  */
 725 static void reset_psn(struct ipath_qp *qp, u32 psn)
 726 {
 727         u32 n = qp->s_last;
 728         struct ipath_swqe *wqe = get_swqe_ptr(qp, n);
 729         u32 opcode;
 730
 731         qp->s_cur = n;
 732
 733         /*
 734          * If we are starting the request from the beginning,
 735          * let the normal send code handle initialization.
 736          */
 737         if (ipath_cmp24(psn, wqe->psn) <= 0) {
 738                 qp->s_state = OP(SEND_LAST);
 739                 goto done;
 740         }
 741
 742         /* Find the work request opcode corresponding to the given PSN. */
 743         opcode = wqe->wr.opcode;
 744         for (;;) {
 745                 int diff;
 746
 747                 if (++n == qp->s_size)
 748                         n = 0;
 749                 if (n == qp->s_tail)
 750                         break;
 751                 wqe = get_swqe_ptr(qp, n);
 752                 diff = ipath_cmp24(psn, wqe->psn);
 753                 if (diff < 0)
 754                         break;
 755                 qp->s_cur = n;
 756                 /*
 757                  * If we are starting the request from the beginning,
 758                  * let the normal send code handle initialization.
 759                  */
 760                 if (diff == 0) {
 761                         qp->s_state = OP(SEND_LAST);
 762                         goto done;
 763                 }
 764                 opcode = wqe->wr.opcode;
 765         }
 766
 767         /*
 768          * Set the state to restart in the middle of a request.
 769          * Don't change the s_sge, s_cur_sge, or s_cur_size.
 770          * See ipath_make_rc_req().
 771          */
 772         switch (opcode) {
 773         case IB_WR_SEND:
 774         case IB_WR_SEND_WITH_IMM:
 775                 qp->s_state = OP(RDMA_READ_RESPONSE_FIRST);
 776                 break;
 777
 778         case IB_WR_RDMA_WRITE:
 779         case IB_WR_RDMA_WRITE_WITH_IMM:
 780                 qp->s_state = OP(RDMA_READ_RESPONSE_LAST);
 781                 break;
 782
 783         case IB_WR_RDMA_READ:
 784                 qp->s_state = OP(RDMA_READ_RESPONSE_MIDDLE);
 785                 break;
 786
 787         default:
 788                 /*
 789                  * This case shouldn't happen since its only
 790                  * one PSN per req.
 791                  */
 792                 qp->s_state = OP(SEND_LAST);
 793         }
 794 done:
 795         qp->s_psn = psn;
 796 }
 797
 798 /**
 799  * ipath_restart_rc - back up requester to resend the last un-ACKed request
 800  * @qp: the QP to restart
 801  * @psn: packet sequence number for the request
 802  * @wc: the work completion request
 803  *
 804  * The QP s_lock should be held and interrupts disabled.
 805  */
 806 void ipath_restart_rc(struct ipath_qp *qp, u32 psn)
 807 {
 808         struct ipath_swqe *wqe = get_swqe_ptr(qp, qp->s_last);
 809         struct ipath_ibdev *dev;
 810
 811         if (qp->s_retry == 0) {
 812                 ipath_send_complete(qp, wqe, IB_WC_RETRY_EXC_ERR);
 813                 ipath_error_qp(qp, IB_WC_WR_FLUSH_ERR);
 814                 goto bail;
 815         }
 816         qp->s_retry--;
 817
 818         /*
 819          * Remove the QP from the timeout queue.
 820          * Note: it may already have been removed by ipath_ib_timer().
 821          */
 822         dev = to_idev(qp->ibqp.device);
 823         spin_lock(&dev->pending_lock);
 824         if (!list_empty(&qp->timerwait))
 825                 list_del_init(&qp->timerwait);
 826         if (!list_empty(&qp->piowait))
 827                 list_del_init(&qp->piowait);
 828         spin_unlock(&dev->pending_lock);
 829
 830         if (wqe->wr.opcode == IB_WR_RDMA_READ)
 831                 dev->n_rc_resends++;
 832         else
 833                 dev->n_rc_resends += (qp->s_psn - psn) & IPATH_PSN_MASK;
 834
 835         reset_psn(qp, psn);
 836         ipath_schedule_send(qp);
 837
 838 bail:
 839         return;
 840 }
 841
 842 static inline void update_last_psn(struct ipath_qp *qp, u32 psn)
 843 {
 844         qp->s_last_psn = psn;
 845 }
 846
 847 /**
 848  * do_rc_ack - process an incoming RC ACK
 849  * @qp: the QP the ACK came in on
 850  * @psn: the packet sequence number of the ACK
 851  * @opcode: the opcode of the request that resulted in the ACK
 852  *
 853  * This is called from ipath_rc_rcv_resp() to process an incoming RC ACK
 854  * for the given QP.
 855  * Called at interrupt level with the QP s_lock held and interrupts disabled.
 856  * Returns 1 if OK, 0 if current operation should be aborted (NAK).
 857  */
 858 static int do_rc_ack(struct ipath_qp *qp, u32 aeth, u32 psn, int opcode,
 859                      u64 val)
 860 {
 861         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
 862         struct ib_wc wc;
 863         enum ib_wc_status status;
 864         struct ipath_swqe *wqe;
 865         int ret = 0;
 866         u32 ack_psn;
 867         int diff;
 868
 869         /*
 870          * Remove the QP from the timeout queue (or RNR timeout queue).
 871          * If ipath_ib_timer() has already removed it,
 872          * it's OK since we hold the QP s_lock and ipath_restart_rc()
 873          * just won't find anything to restart if we ACK everything.
 874          */
 875         spin_lock(&dev->pending_lock);
 876         if (!list_empty(&qp->timerwait))
 877                 list_del_init(&qp->timerwait);
 878         spin_unlock(&dev->pending_lock);
 879
 880         /*
 881          * Note that NAKs implicitly ACK outstanding SEND and RDMA write
 882          * requests and implicitly NAK RDMA read and atomic requests issued
 883          * before the NAK'ed request.  The MSN won't include the NAK'ed
 884          * request but will include an ACK'ed request(s).
 885          */
 886         ack_psn = psn;
 887         if (aeth >> 29)
 888                 ack_psn--;
 889         wqe = get_swqe_ptr(qp, qp->s_last);
 890
 891         /*
 892          * The MSN might be for a later WQE than the PSN indicates so
 893          * only complete WQEs that the PSN finishes.
 894          */
 895         while ((diff = ipath_cmp24(ack_psn, wqe->lpsn)) >= 0) {
 896                 /*
 897                  * RDMA_READ_RESPONSE_ONLY is a special case since
 898                  * we want to generate completion events for everything
 899                  * before the RDMA read, copy the data, then generate
 900                  * the completion for the read.
 901                  */
 902                 if (wqe->wr.opcode == IB_WR_RDMA_READ &&
 903                     opcode == OP(RDMA_READ_RESPONSE_ONLY) &&
 904                     diff == 0) {
 905                         ret = 1;
 906                         goto bail;
 907                 }
 908                 /*
 909                  * If this request is a RDMA read or atomic, and the ACK is
 910                  * for a later operation, this ACK NAKs the RDMA read or
 911                  * atomic.  In other words, only a RDMA_READ_LAST or ONLY
 912                  * can ACK a RDMA read and likewise for atomic ops.  Note
 913                  * that the NAK case can only happen if relaxed ordering is
 914                  * used and requests are sent after an RDMA read or atomic
 915                  * is sent but before the response is received.
 916                  */
 917                 if ((wqe->wr.opcode == IB_WR_RDMA_READ &&
 918                      (opcode != OP(RDMA_READ_RESPONSE_LAST) || diff != 0)) ||
 919                     ((wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
 920                       wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) &&
 921                      (opcode != OP(ATOMIC_ACKNOWLEDGE) || diff != 0))) {
 922                         /*
 923                          * The last valid PSN seen is the previous
 924                          * request's.
 925                          */
 926                         update_last_psn(qp, wqe->psn - 1);
 927                         /* Retry this request. */
 928                         ipath_restart_rc(qp, wqe->psn);
 929                         /*
 930                          * No need to process the ACK/NAK since we are
 931                          * restarting an earlier request.
 932                          */
 933                         goto bail;
 934                 }
 935                 if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
 936                     wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)
 937                         *(u64 *) wqe->sg_list[0].vaddr = val;
 938                 if (qp->s_num_rd_atomic &&
 939                     (wqe->wr.opcode == IB_WR_RDMA_READ ||
 940                      wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
 941                      wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)) {
 942                         qp->s_num_rd_atomic--;
 943                         /* Restart sending task if fence is complete */
 944                         if (((qp->s_flags & IPATH_S_FENCE_PENDING) &&
 945                              !qp->s_num_rd_atomic) ||
 946                             qp->s_flags & IPATH_S_RDMAR_PENDING)
 947                                 ipath_schedule_send(qp);
 948                 }
 949                 /* Post a send completion queue entry if requested. */
 950                 if (!(qp->s_flags & IPATH_S_SIGNAL_REQ_WR) ||
 951                     (wqe->wr.send_flags & IB_SEND_SIGNALED)) {
 952                         memset(&wc, 0, sizeof wc);
 953                         wc.wr_id = wqe->wr.wr_id;
 954                         wc.status = IB_WC_SUCCESS;
 955                         wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
 956                         wc.byte_len = wqe->length;
 957                         wc.qp = &qp->ibqp;
 958                         wc.src_qp = qp->remote_qpn;
 959                         wc.slid = qp->remote_ah_attr.dlid;
 960                         wc.sl = qp->remote_ah_attr.sl;
 961                         ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 0);
 962                 }
 963                 qp->s_retry = qp->s_retry_cnt;
 964                 /*
 965                  * If we are completing a request which is in the process of
 966                  * being resent, we can stop resending it since we know the
 967                  * responder has already seen it.
 968                  */
 969                 if (qp->s_last == qp->s_cur) {
 970                         if (++qp->s_cur >= qp->s_size)
 971                                 qp->s_cur = 0;
 972                         qp->s_last = qp->s_cur;
 973                         if (qp->s_last == qp->s_tail)
 974                                 break;
 975                         wqe = get_swqe_ptr(qp, qp->s_cur);
 976                         qp->s_state = OP(SEND_LAST);
 977                         qp->s_psn = wqe->psn;
 978                 } else {
 979                         if (++qp->s_last >= qp->s_size)
 980                                 qp->s_last = 0;
 981                         if (qp->state == IB_QPS_SQD && qp->s_last == qp->s_cur)
 982                                 qp->s_draining = 0;
 983                         if (qp->s_last == qp->s_tail)
 984                                 break;
 985                         wqe = get_swqe_ptr(qp, qp->s_last);
 986                 }
 987         }
 988
 989         switch (aeth >> 29) {
 990         case 0:         /* ACK */
 991                 dev->n_rc_acks++;
 992                 /* If this is a partial ACK, reset the retransmit timer. */
 993                 if (qp->s_last != qp->s_tail) {
 994                         spin_lock(&dev->pending_lock);
 995                         if (list_empty(&qp->timerwait))
 996                                 list_add_tail(&qp->timerwait,
 997                                         &dev->pending[dev->pending_index]);
 998                         spin_unlock(&dev->pending_lock);
 999                         /*
1000                          * If we get a partial ACK for a resent operation,
1001                          * we can stop resending the earlier packets and
1002                          * continue with the next packet the receiver wants.
1003                          */
1004                         if (ipath_cmp24(qp->s_psn, psn) <= 0) {
1005                                 reset_psn(qp, psn + 1);
1006                                 ipath_schedule_send(qp);
1007                         }
1008                 } else if (ipath_cmp24(qp->s_psn, psn) <= 0) {
1009                         qp->s_state = OP(SEND_LAST);
1010                         qp->s_psn = psn + 1;
1011                 }
1012                 ipath_get_credit(qp, aeth);
1013                 qp->s_rnr_retry = qp->s_rnr_retry_cnt;
1014                 qp->s_retry = qp->s_retry_cnt;
1015                 update_last_psn(qp, psn);
1016                 ret = 1;
1017                 goto bail;
1018
1019         case 1:         /* RNR NAK */
1020                 dev->n_rnr_naks++;
1021                 if (qp->s_last == qp->s_tail)
1022                         goto bail;
1023                 if (qp->s_rnr_retry == 0) {
1024                         status = IB_WC_RNR_RETRY_EXC_ERR;
1025                         goto class_b;
1026                 }
1027                 if (qp->s_rnr_retry_cnt < 7)
1028                         qp->s_rnr_retry--;
1029
1030                 /* The last valid PSN is the previous PSN. */
1031                 update_last_psn(qp, psn - 1);
1032
1033                 if (wqe->wr.opcode == IB_WR_RDMA_READ)
1034                         dev->n_rc_resends++;
1035                 else
1036                         dev->n_rc_resends +=
1037                                 (qp->s_psn - psn) & IPATH_PSN_MASK;
1038
1039                 reset_psn(qp, psn);
1040
1041                 qp->s_rnr_timeout =
1042                         ib_ipath_rnr_table[(aeth >> IPATH_AETH_CREDIT_SHIFT) &
1043                                            IPATH_AETH_CREDIT_MASK];
1044                 ipath_insert_rnr_queue(qp);
1045                 ipath_schedule_send(qp);
1046                 goto bail;
1047
1048         case 3:         /* NAK */
1049                 if (qp->s_last == qp->s_tail)
1050                         goto bail;
1051                 /* The last valid PSN is the previous PSN. */
1052                 update_last_psn(qp, psn - 1);
1053                 switch ((aeth >> IPATH_AETH_CREDIT_SHIFT) &
1054                         IPATH_AETH_CREDIT_MASK) {
1055                 case 0: /* PSN sequence error */
1056                         dev->n_seq_naks++;
1057                         /*
1058                          * Back up to the responder's expected PSN.
1059                          * Note that we might get a NAK in the middle of an
1060                          * RDMA READ response which terminates the RDMA
1061                          * READ.
1062                          */
1063                         ipath_restart_rc(qp, psn);
1064                         break;
1065
1066                 case 1: /* Invalid Request */
1067                         status = IB_WC_REM_INV_REQ_ERR;
1068                         dev->n_other_naks++;
1069                         goto class_b;
1070
1071                 case 2: /* Remote Access Error */
1072                         status = IB_WC_REM_ACCESS_ERR;
1073                         dev->n_other_naks++;
1074                         goto class_b;
1075
1076                 case 3: /* Remote Operation Error */
1077                         status = IB_WC_REM_OP_ERR;
1078                         dev->n_other_naks++;
1079                 class_b:
1080                         ipath_send_complete(qp, wqe, status);
1081                         ipath_error_qp(qp, IB_WC_WR_FLUSH_ERR);
1082                         break;
1083
1084                 default:
1085                         /* Ignore other reserved NAK error codes */
1086                         goto reserved;
1087                 }
1088                 qp->s_rnr_retry = qp->s_rnr_retry_cnt;
1089                 goto bail;
1090
1091         default:                /* 2: reserved */
1092         reserved:
1093                 /* Ignore reserved NAK codes. */
1094                 goto bail;
1095         }
1096
1097 bail:
1098         return ret;
1099 }
1100
1101 /**
1102  * ipath_rc_rcv_resp - process an incoming RC response packet
1103  * @dev: the device this packet came in on
1104  * @ohdr: the other headers for this packet
1105  * @data: the packet data
1106  * @tlen: the packet length
1107  * @qp: the QP for this packet
1108  * @opcode: the opcode for this packet
1109  * @psn: the packet sequence number for this packet
1110  * @hdrsize: the header length
1111  * @pmtu: the path MTU
1112  * @header_in_data: true if part of the header data is in the data buffer
1113  *
1114  * This is called from ipath_rc_rcv() to process an incoming RC response
1115  * packet for the given QP.
1116  * Called at interrupt level.
1117  */
1118 static inline void ipath_rc_rcv_resp(struct ipath_ibdev *dev,
1119                                      struct ipath_other_headers *ohdr,
1120                                      void *data, u32 tlen,
1121                                      struct ipath_qp *qp,
1122                                      u32 opcode,
1123                                      u32 psn, u32 hdrsize, u32 pmtu,
1124                                      int header_in_data)
1125 {
1126         struct ipath_swqe *wqe;
1127         enum ib_wc_status status;
1128         unsigned long flags;
1129         int diff;
1130         u32 pad;
1131         u32 aeth;
1132         u64 val;
1133
1134         spin_lock_irqsave(&qp->s_lock, flags);
1135
1136         /* Double check we can process this now that we hold the s_lock. */
1137         if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_RECV_OK))
1138                 goto ack_done;
1139
1140         /* Ignore invalid responses. */
1141         if (ipath_cmp24(psn, qp->s_next_psn) >= 0)
1142                 goto ack_done;
1143
1144         /* Ignore duplicate responses. */
1145         diff = ipath_cmp24(psn, qp->s_last_psn);
1146         if (unlikely(diff <= 0)) {
1147                 /* Update credits for "ghost" ACKs */
1148                 if (diff == 0 && opcode == OP(ACKNOWLEDGE)) {
1149                         if (!header_in_data)
1150                                 aeth = be32_to_cpu(ohdr->u.aeth);
1151                         else {
1152                                 aeth = be32_to_cpu(((__be32 *) data)[0]);
1153                                 data += sizeof(__be32);
1154                         }
1155                         if ((aeth >> 29) == 0)
1156                                 ipath_get_credit(qp, aeth);
1157                 }
1158                 goto ack_done;
1159         }
1160
1161         if (unlikely(qp->s_last == qp->s_tail))
1162                 goto ack_done;
1163         wqe = get_swqe_ptr(qp, qp->s_last);
1164         status = IB_WC_SUCCESS;
1165
1166         switch (opcode) {
1167         case OP(ACKNOWLEDGE):
1168         case OP(ATOMIC_ACKNOWLEDGE):
1169         case OP(RDMA_READ_RESPONSE_FIRST):
1170                 if (!header_in_data)
1171                         aeth = be32_to_cpu(ohdr->u.aeth);
1172                 else {
1173                         aeth = be32_to_cpu(((__be32 *) data)[0]);
1174                         data += sizeof(__be32);
1175                 }
1176                 if (opcode == OP(ATOMIC_ACKNOWLEDGE)) {
1177                         if (!header_in_data) {
1178                                 __be32 *p = ohdr->u.at.atomic_ack_eth;
1179
1180                                 val = ((u64) be32_to_cpu(p[0]) << 32) |
1181                                         be32_to_cpu(p[1]);
1182                         } else
1183                                 val = be64_to_cpu(((__be64 *) data)[0]);
1184                 } else
1185                         val = 0;
1186                 if (!do_rc_ack(qp, aeth, psn, opcode, val) ||
1187                     opcode != OP(RDMA_READ_RESPONSE_FIRST))
1188                         goto ack_done;
1189                 hdrsize += 4;
1190                 wqe = get_swqe_ptr(qp, qp->s_last);
1191                 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1192                         goto ack_op_err;
1193                 qp->r_flags &= ~IPATH_R_RDMAR_SEQ;
1194                 /*
1195                  * If this is a response to a resent RDMA read, we
1196                  * have to be careful to copy the data to the right
1197                  * location.
1198                  */
1199                 qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
1200                                                   wqe, psn, pmtu);
1201                 goto read_middle;
1202
1203         case OP(RDMA_READ_RESPONSE_MIDDLE):
1204                 /* no AETH, no ACK */
1205                 if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) {
1206                         dev->n_rdma_seq++;
1207                         if (qp->r_flags & IPATH_R_RDMAR_SEQ)
1208                                 goto ack_done;
1209                         qp->r_flags |= IPATH_R_RDMAR_SEQ;
1210                         ipath_restart_rc(qp, qp->s_last_psn + 1);
1211                         goto ack_done;
1212                 }
1213                 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1214                         goto ack_op_err;
1215         read_middle:
1216                 if (unlikely(tlen != (hdrsize + pmtu + 4)))
1217                         goto ack_len_err;
1218                 if (unlikely(pmtu >= qp->s_rdma_read_len))
1219                         goto ack_len_err;
1220
1221                 /* We got a response so update the timeout. */
1222                 spin_lock(&dev->pending_lock);
1223                 if (qp->s_rnr_timeout == 0 && !list_empty(&qp->timerwait))
1224                         list_move_tail(&qp->timerwait,
1225                                        &dev->pending[dev->pending_index]);
1226                 spin_unlock(&dev->pending_lock);
1227
1228                 if (opcode == OP(RDMA_READ_RESPONSE_MIDDLE))
1229                         qp->s_retry = qp->s_retry_cnt;
1230
1231                 /*
1232                  * Update the RDMA receive state but do the copy w/o
1233                  * holding the locks and blocking interrupts.
1234                  */
1235                 qp->s_rdma_read_len -= pmtu;
1236                 update_last_psn(qp, psn);
1237                 spin_unlock_irqrestore(&qp->s_lock, flags);
1238                 ipath_copy_sge(&qp->s_rdma_read_sge, data, pmtu);
1239                 goto bail;
1240
1241         case OP(RDMA_READ_RESPONSE_ONLY):
1242                 if (!header_in_data)
1243                         aeth = be32_to_cpu(ohdr->u.aeth);
1244                 else
1245                         aeth = be32_to_cpu(((__be32 *) data)[0]);
1246                 if (!do_rc_ack(qp, aeth, psn, opcode, 0))
1247                         goto ack_done;
1248                 /* Get the number of bytes the message was padded by. */
1249                 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1250                 /*
1251                  * Check that the data size is >= 0 && <= pmtu.
1252                  * Remember to account for the AETH header (4) and
1253                  * ICRC (4).
1254                  */
1255                 if (unlikely(tlen < (hdrsize + pad + 8)))
1256                         goto ack_len_err;
1257                 /*
1258                  * If this is a response to a resent RDMA read, we
1259                  * have to be careful to copy the data to the right
1260                  * location.
1261                  */
1262                 wqe = get_swqe_ptr(qp, qp->s_last);
1263                 qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
1264                                                   wqe, psn, pmtu);
1265                 goto read_last;
1266
1267         case OP(RDMA_READ_RESPONSE_LAST):
1268                 /* ACKs READ req. */
1269                 if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) {
1270                         dev->n_rdma_seq++;
1271                         if (qp->r_flags & IPATH_R_RDMAR_SEQ)
1272                                 goto ack_done;
1273                         qp->r_flags |= IPATH_R_RDMAR_SEQ;
1274                         ipath_restart_rc(qp, qp->s_last_psn + 1);
1275                         goto ack_done;
1276                 }
1277                 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1278                         goto ack_op_err;
1279                 /* Get the number of bytes the message was padded by. */
1280                 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1281                 /*
1282                  * Check that the data size is >= 1 && <= pmtu.
1283                  * Remember to account for the AETH header (4) and
1284                  * ICRC (4).
1285                  */
1286                 if (unlikely(tlen <= (hdrsize + pad + 8)))
1287                         goto ack_len_err;
1288         read_last:
1289                 tlen -= hdrsize + pad + 8;
1290                 if (unlikely(tlen != qp->s_rdma_read_len))
1291                         goto ack_len_err;
1292                 if (!header_in_data)
1293                         aeth = be32_to_cpu(ohdr->u.aeth);
1294                 else {
1295                         aeth = be32_to_cpu(((__be32 *) data)[0]);
1296                         data += sizeof(__be32);
1297                 }
1298                 ipath_copy_sge(&qp->s_rdma_read_sge, data, tlen);
1299                 (void) do_rc_ack(qp, aeth, psn,
1300                                  OP(RDMA_READ_RESPONSE_LAST), 0);
1301                 goto ack_done;
1302         }
1303
1304 ack_op_err:
1305         status = IB_WC_LOC_QP_OP_ERR;
1306         goto ack_err;
1307
1308 ack_len_err:
1309         status = IB_WC_LOC_LEN_ERR;
1310 ack_err:
1311         ipath_send_complete(qp, wqe, status);
1312         ipath_error_qp(qp, IB_WC_WR_FLUSH_ERR);
1313 ack_done:
1314         spin_unlock_irqrestore(&qp->s_lock, flags);
1315 bail:
1316         return;
1317 }
1318
1319 /**
1320  * ipath_rc_rcv_error - process an incoming duplicate or error RC packet
1321  * @dev: the device this packet came in on
1322  * @ohdr: the other headers for this packet
1323  * @data: the packet data
1324  * @qp: the QP for this packet
1325  * @opcode: the opcode for this packet
1326  * @psn: the packet sequence number for this packet
1327  * @diff: the difference between the PSN and the expected PSN
1328  * @header_in_data: true if part of the header data is in the data buffer
1329  *
1330  * This is called from ipath_rc_rcv() to process an unexpected
1331  * incoming RC packet for the given QP.
1332  * Called at interrupt level.
1333  * Return 1 if no more processing is needed; otherwise return 0 to
1334  * schedule a response to be sent.
1335  */
1336 static inline int ipath_rc_rcv_error(struct ipath_ibdev *dev,
1337                                      struct ipath_other_headers *ohdr,
1338                                      void *data,
1339                                      struct ipath_qp *qp,
1340                                      u32 opcode,
1341                                      u32 psn,
1342                                      int diff,
1343                                      int header_in_data)
1344 {
1345         struct ipath_ack_entry *e;
1346         u8 i, prev;
1347         int old_req;
1348         unsigned long flags;
1349
1350         if (diff > 0) {
1351                 /*
1352                  * Packet sequence error.
1353                  * A NAK will ACK earlier sends and RDMA writes.
1354                  * Don't queue the NAK if we already sent one.
1355                  */
1356                 if (!qp->r_nak_state) {
1357                         qp->r_nak_state = IB_NAK_PSN_ERROR;
1358                         /* Use the expected PSN. */
1359                         qp->r_ack_psn = qp->r_psn;
1360                         goto send_ack;
1361                 }
1362                 goto done;
1363         }
1364
1365         /*
1366          * Handle a duplicate request.  Don't re-execute SEND, RDMA
1367          * write or atomic op.  Don't NAK errors, just silently drop
1368          * the duplicate request.  Note that r_sge, r_len, and
1369          * r_rcv_len may be in use so don't modify them.
1370          *
1371          * We are supposed to ACK the earliest duplicate PSN but we
1372          * can coalesce an outstanding duplicate ACK.  We have to
1373          * send the earliest so that RDMA reads can be restarted at
1374          * the requester's expected PSN.
1375          *
1376          * First, find where this duplicate PSN falls within the
1377          * ACKs previously sent.
1378          */
1379         psn &= IPATH_PSN_MASK;
1380         e = NULL;
1381         old_req = 1;
1382
1383         spin_lock_irqsave(&qp->s_lock, flags);
1384         /* Double check we can process this now that we hold the s_lock. */
1385         if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_RECV_OK))
1386                 goto unlock_done;
1387
1388         for (i = qp->r_head_ack_queue; ; i = prev) {
1389                 if (i == qp->s_tail_ack_queue)
1390                         old_req = 0;
1391                 if (i)
1392                         prev = i - 1;
1393                 else
1394                         prev = IPATH_MAX_RDMA_ATOMIC;
1395                 if (prev == qp->r_head_ack_queue) {
1396                         e = NULL;
1397                         break;
1398                 }
1399                 e = &qp->s_ack_queue[prev];
1400                 if (!e->opcode) {
1401                         e = NULL;
1402                         break;
1403                 }
1404                 if (ipath_cmp24(psn, e->psn) >= 0) {
1405                         if (prev == qp->s_tail_ack_queue)
1406                                 old_req = 0;
1407                         break;
1408                 }
1409         }
1410         switch (opcode) {
1411         case OP(RDMA_READ_REQUEST): {
1412                 struct ib_reth *reth;
1413                 u32 offset;
1414                 u32 len;
1415
1416                 /*
1417                  * If we didn't find the RDMA read request in the ack queue,
1418                  * or the send tasklet is already backed up to send an
1419                  * earlier entry, we can ignore this request.
1420                  */
1421                 if (!e || e->opcode != OP(RDMA_READ_REQUEST) || old_req)
1422                         goto unlock_done;
1423                 /* RETH comes after BTH */
1424                 if (!header_in_data)
1425                         reth = &ohdr->u.rc.reth;
1426                 else {
1427                         reth = (struct ib_reth *)data;
1428                         data += sizeof(*reth);
1429                 }
1430                 /*
1431                  * Address range must be a subset of the original
1432                  * request and start on pmtu boundaries.
1433                  * We reuse the old ack_queue slot since the requester
1434                  * should not back up and request an earlier PSN for the
1435                  * same request.
1436                  */
1437                 offset = ((psn - e->psn) & IPATH_PSN_MASK) *
1438                         ib_mtu_enum_to_int(qp->path_mtu);
1439                 len = be32_to_cpu(reth->length);
1440                 if (unlikely(offset + len > e->rdma_sge.sge.sge_length))
1441                         goto unlock_done;
1442                 if (len != 0) {
1443                         u32 rkey = be32_to_cpu(reth->rkey);
1444                         u64 vaddr = be64_to_cpu(reth->vaddr);
1445                         int ok;
1446
1447                         ok = ipath_rkey_ok(qp, &e->rdma_sge,
1448                                            len, vaddr, rkey,
1449                                            IB_ACCESS_REMOTE_READ);
1450                         if (unlikely(!ok))
1451                                 goto unlock_done;
1452                 } else {
1453                         e->rdma_sge.sg_list = NULL;
1454                         e->rdma_sge.num_sge = 0;
1455                         e->rdma_sge.sge.mr = NULL;
1456                         e->rdma_sge.sge.vaddr = NULL;
1457                         e->rdma_sge.sge.length = 0;
1458                         e->rdma_sge.sge.sge_length = 0;
1459                 }
1460                 e->psn = psn;
1461                 qp->s_ack_state = OP(ACKNOWLEDGE);
1462                 qp->s_tail_ack_queue = prev;
1463                 break;
1464         }
1465
1466         case OP(COMPARE_SWAP):
1467         case OP(FETCH_ADD): {
1468                 /*
1469                  * If we didn't find the atomic request in the ack queue
1470                  * or the send tasklet is already backed up to send an
1471                  * earlier entry, we can ignore this request.
1472                  */
1473                 if (!e || e->opcode != (u8) opcode || old_req)
1474                         goto unlock_done;
1475                 qp->s_ack_state = OP(ACKNOWLEDGE);
1476                 qp->s_tail_ack_queue = prev;
1477                 break;
1478         }
1479
1480         default:
1481                 if (old_req)
1482                         goto unlock_done;
1483                 /*
1484                  * Resend the most recent ACK if this request is
1485                  * after all the previous RDMA reads and atomics.
1486                  */
1487                 if (i == qp->r_head_ack_queue) {
1488                         spin_unlock_irqrestore(&qp->s_lock, flags);
1489                         qp->r_nak_state = 0;
1490                         qp->r_ack_psn = qp->r_psn - 1;
1491                         goto send_ack;
1492                 }
1493                 /*
1494                  * Try to send a simple ACK to work around a Mellanox bug
1495                  * which doesn't accept a RDMA read response or atomic
1496                  * response as an ACK for earlier SENDs or RDMA writes.
1497                  */
1498                 if (qp->r_head_ack_queue == qp->s_tail_ack_queue &&
1499                     !(qp->s_flags & IPATH_S_ACK_PENDING) &&
1500                     qp->s_ack_state == OP(ACKNOWLEDGE)) {
1501                         spin_unlock_irqrestore(&qp->s_lock, flags);
1502                         qp->r_nak_state = 0;
1503                         qp->r_ack_psn = qp->s_ack_queue[i].psn - 1;
1504                         goto send_ack;
1505                 }
1506                 /*
1507                  * Resend the RDMA read or atomic op which
1508                  * ACKs this duplicate request.
1509                  */
1510                 qp->s_ack_state = OP(ACKNOWLEDGE);
1511                 qp->s_tail_ack_queue = i;
1512                 break;
1513         }
1514         qp->r_nak_state = 0;
1515         ipath_schedule_send(qp);
1516
1517 unlock_done:
1518         spin_unlock_irqrestore(&qp->s_lock, flags);
1519 done:
1520         return 1;
1521
1522 send_ack:
1523         return 0;
1524 }
1525
1526 void ipath_rc_error(struct ipath_qp *qp, enum ib_wc_status err)
1527 {
1528         unsigned long flags;
1529         int lastwqe;
1530
1531         spin_lock_irqsave(&qp->s_lock, flags);
1532         lastwqe = ipath_error_qp(qp, err);
1533         spin_unlock_irqrestore(&qp->s_lock, flags);
1534
1535         if (lastwqe) {
1536                 struct ib_event ev;
1537
1538                 ev.device = qp->ibqp.device;
1539                 ev.element.qp = &qp->ibqp;
1540                 ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
1541                 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
1542         }
1543 }
1544
1545 static inline void ipath_update_ack_queue(struct ipath_qp *qp, unsigned n)
1546 {
1547         unsigned next;
1548
1549         next = n + 1;
1550         if (next > IPATH_MAX_RDMA_ATOMIC)
1551                 next = 0;
1552         if (n == qp->s_tail_ack_queue) {
1553                 qp->s_tail_ack_queue = next;
1554                 qp->s_ack_state = OP(ACKNOWLEDGE);
1555         }
1556 }
1557
1558 /**
1559  * ipath_rc_rcv - process an incoming RC packet
1560  * @dev: the device this packet came in on
1561  * @hdr: the header of this packet
1562  * @has_grh: true if the header has a GRH
1563  * @data: the packet data
1564  * @tlen: the packet length
1565  * @qp: the QP for this packet
1566  *
1567  * This is called from ipath_qp_rcv() to process an incoming RC packet
1568  * for the given QP.
1569  * Called at interrupt level.
1570  */
1571 void ipath_rc_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
1572                   int has_grh, void *data, u32 tlen, struct ipath_qp *qp)
1573 {
1574         struct ipath_other_headers *ohdr;
1575         u32 opcode;
1576         u32 hdrsize;
1577         u32 psn;
1578         u32 pad;
1579         struct ib_wc wc;
1580         u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
1581         int diff;
1582         struct ib_reth *reth;
1583         int header_in_data;
1584         unsigned long flags;
1585
1586         /* Validate the SLID. See Ch. 9.6.1.5 */
1587         if (unlikely(be16_to_cpu(hdr->lrh[3]) != qp->remote_ah_attr.dlid))
1588                 goto done;
1589
1590         /* Check for GRH */
1591         if (!has_grh) {
1592                 ohdr = &hdr->u.oth;
1593                 hdrsize = 8 + 12;       /* LRH + BTH */
1594                 psn = be32_to_cpu(ohdr->bth[2]);
1595                 header_in_data = 0;
1596         } else {
1597                 ohdr = &hdr->u.l.oth;
1598                 hdrsize = 8 + 40 + 12;  /* LRH + GRH + BTH */
1599                 /*
1600                  * The header with GRH is 60 bytes and the core driver sets
1601                  * the eager header buffer size to 56 bytes so the last 4
1602                  * bytes of the BTH header (PSN) is in the data buffer.
1603                  */
1604                 header_in_data = dev->dd->ipath_rcvhdrentsize == 16;
1605                 if (header_in_data) {
1606                         psn = be32_to_cpu(((__be32 *) data)[0]);
1607                         data += sizeof(__be32);
1608                 } else
1609                         psn = be32_to_cpu(ohdr->bth[2]);
1610         }
1611
1612         /*
1613          * Process responses (ACKs) before anything else.  Note that the
1614          * packet sequence number will be for something in the send work
1615          * queue rather than the expected receive packet sequence number.
1616          * In other words, this QP is the requester.
1617          */
1618         opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
1619         if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
1620             opcode <= OP(ATOMIC_ACKNOWLEDGE)) {
1621                 ipath_rc_rcv_resp(dev, ohdr, data, tlen, qp, opcode, psn,
1622                                   hdrsize, pmtu, header_in_data);
1623                 goto done;
1624         }
1625
1626         /* Compute 24 bits worth of difference. */
1627         diff = ipath_cmp24(psn, qp->r_psn);
1628         if (unlikely(diff)) {
1629                 if (ipath_rc_rcv_error(dev, ohdr, data, qp, opcode,
1630                                        psn, diff, header_in_data))
1631                         goto done;
1632                 goto send_ack;
1633         }
1634
1635         /* Check for opcode sequence errors. */
1636         switch (qp->r_state) {
1637         case OP(SEND_FIRST):
1638         case OP(SEND_MIDDLE):
1639                 if (opcode == OP(SEND_MIDDLE) ||
1640                     opcode == OP(SEND_LAST) ||
1641                     opcode == OP(SEND_LAST_WITH_IMMEDIATE))
1642                         break;
1643                 goto nack_inv;
1644
1645         case OP(RDMA_WRITE_FIRST):
1646         case OP(RDMA_WRITE_MIDDLE):
1647                 if (opcode == OP(RDMA_WRITE_MIDDLE) ||
1648                     opcode == OP(RDMA_WRITE_LAST) ||
1649                     opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
1650                         break;
1651                 goto nack_inv;
1652
1653         default:
1654                 if (opcode == OP(SEND_MIDDLE) ||
1655                     opcode == OP(SEND_LAST) ||
1656                     opcode == OP(SEND_LAST_WITH_IMMEDIATE) ||
1657                     opcode == OP(RDMA_WRITE_MIDDLE) ||
1658                     opcode == OP(RDMA_WRITE_LAST) ||
1659                     opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
1660                         goto nack_inv;
1661                 /*
1662                  * Note that it is up to the requester to not send a new
1663                  * RDMA read or atomic operation before receiving an ACK
1664                  * for the previous operation.
1665                  */
1666                 break;
1667         }
1668
1669         memset(&wc, 0, sizeof wc);
1670
1671         /* OK, process the packet. */
1672         switch (opcode) {
1673         case OP(SEND_FIRST):
1674                 if (!ipath_get_rwqe(qp, 0))
1675                         goto rnr_nak;
1676                 qp->r_rcv_len = 0;
1677                 /* FALLTHROUGH */
1678         case OP(SEND_MIDDLE):
1679         case OP(RDMA_WRITE_MIDDLE):
1680         send_middle:
1681                 /* Check for invalid length PMTU or posted rwqe len. */
1682                 if (unlikely(tlen != (hdrsize + pmtu + 4)))
1683                         goto nack_inv;
1684                 qp->r_rcv_len += pmtu;
1685                 if (unlikely(qp->r_rcv_len > qp->r_len))
1686                         goto nack_inv;
1687                 ipath_copy_sge(&qp->r_sge, data, pmtu);
1688                 break;
1689
1690         case OP(RDMA_WRITE_LAST_WITH_IMMEDIATE):
1691                 /* consume RWQE */
1692                 if (!ipath_get_rwqe(qp, 1))
1693                         goto rnr_nak;
1694                 goto send_last_imm;
1695
1696         case OP(SEND_ONLY):
1697         case OP(SEND_ONLY_WITH_IMMEDIATE):
1698                 if (!ipath_get_rwqe(qp, 0))
1699                         goto rnr_nak;
1700                 qp->r_rcv_len = 0;
1701                 if (opcode == OP(SEND_ONLY))
1702                         goto send_last;
1703                 /* FALLTHROUGH */
1704         case OP(SEND_LAST_WITH_IMMEDIATE):
1705         send_last_imm:
1706                 if (header_in_data) {
1707                         wc.ex.imm_data = *(__be32 *) data;
1708                         data += sizeof(__be32);
1709                 } else {
1710                         /* Immediate data comes after BTH */
1711                         wc.ex.imm_data = ohdr->u.imm_data;
1712                 }
1713                 hdrsize += 4;
1714                 wc.wc_flags = IB_WC_WITH_IMM;
1715                 /* FALLTHROUGH */
1716         case OP(SEND_LAST):
1717         case OP(RDMA_WRITE_LAST):
1718         send_last:
1719                 /* Get the number of bytes the message was padded by. */
1720                 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1721                 /* Check for invalid length. */
1722                 /* XXX LAST len should be >= 1 */
1723                 if (unlikely(tlen < (hdrsize + pad + 4)))
1724                         goto nack_inv;
1725                 /* Don't count the CRC. */
1726                 tlen -= (hdrsize + pad + 4);
1727                 wc.byte_len = tlen + qp->r_rcv_len;
1728                 if (unlikely(wc.byte_len > qp->r_len))
1729                         goto nack_inv;
1730                 ipath_copy_sge(&qp->r_sge, data, tlen);
1731                 qp->r_msn++;
1732                 if (!test_and_clear_bit(IPATH_R_WRID_VALID, &qp->r_aflags))
1733                         break;
1734                 wc.wr_id = qp->r_wr_id;
1735                 wc.status = IB_WC_SUCCESS;
1736                 if (opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE) ||
1737                     opcode == OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE))
1738                         wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
1739                 else
1740                         wc.opcode = IB_WC_RECV;
1741                 wc.qp = &qp->ibqp;
1742                 wc.src_qp = qp->remote_qpn;
1743                 wc.slid = qp->remote_ah_attr.dlid;
1744                 wc.sl = qp->remote_ah_attr.sl;
1745                 /* Signal completion event if the solicited bit is set. */
1746                 ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
1747                                (ohdr->bth[0] &
1748                                 __constant_cpu_to_be32(1 << 23)) != 0);
1749                 break;
1750
1751         case OP(RDMA_WRITE_FIRST):
1752         case OP(RDMA_WRITE_ONLY):
1753         case OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE):
1754                 if (unlikely(!(qp->qp_access_flags &
1755                                IB_ACCESS_REMOTE_WRITE)))
1756                         goto nack_inv;
1757                 /* consume RWQE */
1758                 /* RETH comes after BTH */
1759                 if (!header_in_data)
1760                         reth = &ohdr->u.rc.reth;
1761                 else {
1762                         reth = (struct ib_reth *)data;
1763                         data += sizeof(*reth);
1764                 }
1765                 hdrsize += sizeof(*reth);
1766                 qp->r_len = be32_to_cpu(reth->length);
1767                 qp->r_rcv_len = 0;
1768                 if (qp->r_len != 0) {
1769                         u32 rkey = be32_to_cpu(reth->rkey);
1770                         u64 vaddr = be64_to_cpu(reth->vaddr);
1771                         int ok;
1772
1773                         /* Check rkey & NAK */
1774                         ok = ipath_rkey_ok(qp, &qp->r_sge,
1775                                            qp->r_len, vaddr, rkey,
1776                                            IB_ACCESS_REMOTE_WRITE);
1777                         if (unlikely(!ok))
1778                                 goto nack_acc;
1779                 } else {
1780                         qp->r_sge.sg_list = NULL;
1781                         qp->r_sge.sge.mr = NULL;
1782                         qp->r_sge.sge.vaddr = NULL;
1783                         qp->r_sge.sge.length = 0;
1784                         qp->r_sge.sge.sge_length = 0;
1785                 }
1786                 if (opcode == OP(RDMA_WRITE_FIRST))
1787                         goto send_middle;
1788                 else if (opcode == OP(RDMA_WRITE_ONLY))
1789                         goto send_last;
1790                 if (!ipath_get_rwqe(qp, 1))
1791                         goto rnr_nak;
1792                 goto send_last_imm;
1793
1794         case OP(RDMA_READ_REQUEST): {
1795                 struct ipath_ack_entry *e;
1796                 u32 len;
1797                 u8 next;
1798
1799                 if (unlikely(!(qp->qp_access_flags &
1800                                IB_ACCESS_REMOTE_READ)))
1801                         goto nack_inv;
1802                 next = qp->r_head_ack_queue + 1;
1803                 if (next > IPATH_MAX_RDMA_ATOMIC)
1804                         next = 0;
1805                 spin_lock_irqsave(&qp->s_lock, flags);
1806                 /* Double check we can process this while holding the s_lock. */
1807                 if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_RECV_OK))
1808                         goto unlock;
1809                 if (unlikely(next == qp->s_tail_ack_queue)) {
1810                         if (!qp->s_ack_queue[next].sent)
1811                                 goto nack_inv_unlck;
1812                         ipath_update_ack_queue(qp, next);
1813                 }
1814                 e = &qp->s_ack_queue[qp->r_head_ack_queue];
1815                 /* RETH comes after BTH */
1816                 if (!header_in_data)
1817                         reth = &ohdr->u.rc.reth;
1818                 else {
1819                         reth = (struct ib_reth *)data;
1820                         data += sizeof(*reth);
1821                 }
1822                 len = be32_to_cpu(reth->length);
1823                 if (len) {
1824                         u32 rkey = be32_to_cpu(reth->rkey);
1825                         u64 vaddr = be64_to_cpu(reth->vaddr);
1826                         int ok;
1827
1828                         /* Check rkey & NAK */
1829                         ok = ipath_rkey_ok(qp, &e->rdma_sge, len, vaddr,
1830                                            rkey, IB_ACCESS_REMOTE_READ);
1831                         if (unlikely(!ok))
1832                                 goto nack_acc_unlck;
1833                         /*
1834                          * Update the next expected PSN.  We add 1 later
1835                          * below, so only add the remainder here.
1836                          */
1837                         if (len > pmtu)
1838                                 qp->r_psn += (len - 1) / pmtu;
1839                 } else {
1840                         e->rdma_sge.sg_list = NULL;
1841                         e->rdma_sge.num_sge = 0;
1842                         e->rdma_sge.sge.mr = NULL;
1843                         e->rdma_sge.sge.vaddr = NULL;
1844                         e->rdma_sge.sge.length = 0;
1845                         e->rdma_sge.sge.sge_length = 0;
1846                 }
1847                 e->opcode = opcode;
1848                 e->sent = 0;
1849                 e->psn = psn;
1850                 /*
1851                  * We need to increment the MSN here instead of when we
1852                  * finish sending the result since a duplicate request would
1853                  * increment it more than once.
1854                  */
1855                 qp->r_msn++;
1856                 qp->r_psn++;
1857                 qp->r_state = opcode;
1858                 qp->r_nak_state = 0;
1859                 qp->r_head_ack_queue = next;
1860
1861                 /* Schedule the send tasklet. */
1862                 ipath_schedule_send(qp);
1863
1864                 goto unlock;
1865         }
1866
1867         case OP(COMPARE_SWAP):
1868         case OP(FETCH_ADD): {
1869                 struct ib_atomic_eth *ateth;
1870                 struct ipath_ack_entry *e;
1871                 u64 vaddr;
1872                 atomic64_t *maddr;
1873                 u64 sdata;
1874                 u32 rkey;
1875                 u8 next;
1876
1877                 if (unlikely(!(qp->qp_access_flags &
1878                                IB_ACCESS_REMOTE_ATOMIC)))
1879                         goto nack_inv;
1880                 next = qp->r_head_ack_queue + 1;
1881                 if (next > IPATH_MAX_RDMA_ATOMIC)
1882                         next = 0;
1883                 spin_lock_irqsave(&qp->s_lock, flags);
1884                 /* Double check we can process this while holding the s_lock. */
1885                 if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_RECV_OK))
1886                         goto unlock;
1887                 if (unlikely(next == qp->s_tail_ack_queue)) {
1888                         if (!qp->s_ack_queue[next].sent)
1889                                 goto nack_inv_unlck;
1890                         ipath_update_ack_queue(qp, next);
1891                 }
1892                 if (!header_in_data)
1893                         ateth = &ohdr->u.atomic_eth;
1894                 else
1895                         ateth = (struct ib_atomic_eth *)data;
1896                 vaddr = ((u64) be32_to_cpu(ateth->vaddr[0]) << 32) |
1897                         be32_to_cpu(ateth->vaddr[1]);
1898                 if (unlikely(vaddr & (sizeof(u64) - 1)))
1899                         goto nack_inv_unlck;
1900                 rkey = be32_to_cpu(ateth->rkey);
1901                 /* Check rkey & NAK */
1902                 if (unlikely(!ipath_rkey_ok(qp, &qp->r_sge,
1903                                             sizeof(u64), vaddr, rkey,
1904                                             IB_ACCESS_REMOTE_ATOMIC)))
1905                         goto nack_acc_unlck;
1906                 /* Perform atomic OP and save result. */
1907                 maddr = (atomic64_t *) qp->r_sge.sge.vaddr;
1908                 sdata = be64_to_cpu(ateth->swap_data);
1909                 e = &qp->s_ack_queue[qp->r_head_ack_queue];
1910                 e->atomic_data = (opcode == OP(FETCH_ADD)) ?
1911                         (u64) atomic64_add_return(sdata, maddr) - sdata :
1912                         (u64) cmpxchg((u64 *) qp->r_sge.sge.vaddr,
1913                                       be64_to_cpu(ateth->compare_data),
1914                                       sdata);
1915                 e->opcode = opcode;
1916                 e->sent = 0;
1917                 e->psn = psn & IPATH_PSN_MASK;
1918                 qp->r_msn++;
1919                 qp->r_psn++;
1920                 qp->r_state = opcode;
1921                 qp->r_nak_state = 0;
1922                 qp->r_head_ack_queue = next;
1923
1924                 /* Schedule the send tasklet. */
1925                 ipath_schedule_send(qp);
1926
1927                 goto unlock;
1928         }
1929
1930         default:
1931                 /* NAK unknown opcodes. */
1932                 goto nack_inv;
1933         }
1934         qp->r_psn++;
1935         qp->r_state = opcode;
1936         qp->r_ack_psn = psn;
1937         qp->r_nak_state = 0;
1938         /* Send an ACK if requested or required. */
1939         if (psn & (1 << 31))
1940                 goto send_ack;
1941         goto done;
1942
1943 rnr_nak:
1944         qp->r_nak_state = IB_RNR_NAK | qp->r_min_rnr_timer;
1945         qp->r_ack_psn = qp->r_psn;
1946         goto send_ack;
1947
1948 nack_inv_unlck:
1949         spin_unlock_irqrestore(&qp->s_lock, flags);
1950 nack_inv:
1951         ipath_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
1952         qp->r_nak_state = IB_NAK_INVALID_REQUEST;
1953         qp->r_ack_psn = qp->r_psn;
1954         goto send_ack;
1955
1956 nack_acc_unlck:
1957         spin_unlock_irqrestore(&qp->s_lock, flags);
1958 nack_acc:
1959         ipath_rc_error(qp, IB_WC_LOC_PROT_ERR);
1960         qp->r_nak_state = IB_NAK_REMOTE_ACCESS_ERROR;
1961         qp->r_ack_psn = qp->r_psn;
1962 send_ack:
1963         send_rc_ack(qp);
1964         goto done;
1965
1966 unlock:
1967         spin_unlock_irqrestore(&qp->s_lock, flags);
1968 done:
1969         return;
1970 }