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.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.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.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.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, struct ib_wc *wc)
 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                 wc->wr_id = wqe->wr.wr_id;
 781                 wc->status = IB_WC_RETRY_EXC_ERR;
 782                 wc->opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
 783                 wc->vendor_err = 0;
 784                 wc->byte_len = 0;
 785                 wc->qp = &qp->ibqp;
 786                 wc->imm_data = 0;
 787                 wc->src_qp = qp->remote_qpn;
 788                 wc->wc_flags = 0;
 789                 wc->pkey_index = 0;
 790                 wc->slid = qp->remote_ah_attr.dlid;
 791                 wc->sl = qp->remote_ah_attr.sl;
 792                 wc->dlid_path_bits = 0;
 793                 wc->port_num = 0;
 794                 ipath_sqerror_qp(qp, wc);
 795                 goto bail;
 796         }
 797         qp->s_retry--;
 798
 799         /*
 800          * Remove the QP from the timeout queue.
 801          * Note: it may already have been removed by ipath_ib_timer().
 802          */
 803         dev = to_idev(qp->ibqp.device);
 804         spin_lock(&dev->pending_lock);
 805         if (!list_empty(&qp->timerwait))
 806                 list_del_init(&qp->timerwait);
 807         spin_unlock(&dev->pending_lock);
 808
 809         if (wqe->wr.opcode == IB_WR_RDMA_READ)
 810                 dev->n_rc_resends++;
 811         else
 812                 dev->n_rc_resends += (qp->s_psn - psn) & IPATH_PSN_MASK;
 813
 814         reset_psn(qp, psn);
 815         tasklet_hi_schedule(&qp->s_task);
 816
 817 bail:
 818         return;
 819 }
 820
 821 static inline void update_last_psn(struct ipath_qp *qp, u32 psn)
 822 {
 823         if (qp->s_last_psn != psn) {
 824                 qp->s_last_psn = psn;
 825                 if (qp->s_wait_credit) {
 826                         qp->s_wait_credit = 0;
 827                         tasklet_hi_schedule(&qp->s_task);
 828                 }
 829         }
 830 }
 831
 832 /**
 833  * do_rc_ack - process an incoming RC ACK
 834  * @qp: the QP the ACK came in on
 835  * @psn: the packet sequence number of the ACK
 836  * @opcode: the opcode of the request that resulted in the ACK
 837  *
 838  * This is called from ipath_rc_rcv_resp() to process an incoming RC ACK
 839  * for the given QP.
 840  * Called at interrupt level with the QP s_lock held and interrupts disabled.
 841  * Returns 1 if OK, 0 if current operation should be aborted (NAK).
 842  */
 843 static int do_rc_ack(struct ipath_qp *qp, u32 aeth, u32 psn, int opcode,
 844                      u64 val)
 845 {
 846         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
 847         struct ib_wc wc;
 848         struct ipath_swqe *wqe;
 849         int ret = 0;
 850         u32 ack_psn;
 851         int diff;
 852
 853         /*
 854          * Remove the QP from the timeout queue (or RNR timeout queue).
 855          * If ipath_ib_timer() has already removed it,
 856          * it's OK since we hold the QP s_lock and ipath_restart_rc()
 857          * just won't find anything to restart if we ACK everything.
 858          */
 859         spin_lock(&dev->pending_lock);
 860         if (!list_empty(&qp->timerwait))
 861                 list_del_init(&qp->timerwait);
 862         spin_unlock(&dev->pending_lock);
 863
 864         /*
 865          * Note that NAKs implicitly ACK outstanding SEND and RDMA write
 866          * requests and implicitly NAK RDMA read and atomic requests issued
 867          * before the NAK'ed request.  The MSN won't include the NAK'ed
 868          * request but will include an ACK'ed request(s).
 869          */
 870         ack_psn = psn;
 871         if (aeth >> 29)
 872                 ack_psn--;
 873         wqe = get_swqe_ptr(qp, qp->s_last);
 874
 875         /*
 876          * The MSN might be for a later WQE than the PSN indicates so
 877          * only complete WQEs that the PSN finishes.
 878          */
 879         while ((diff = ipath_cmp24(ack_psn, wqe->lpsn)) >= 0) {
 880                 /*
 881                  * RDMA_READ_RESPONSE_ONLY is a special case since
 882                  * we want to generate completion events for everything
 883                  * before the RDMA read, copy the data, then generate
 884                  * the completion for the read.
 885                  */
 886                 if (wqe->wr.opcode == IB_WR_RDMA_READ &&
 887                     opcode == OP(RDMA_READ_RESPONSE_ONLY) &&
 888                     diff == 0) {
 889                         ret = 1;
 890                         goto bail;
 891                 }
 892                 /*
 893                  * If this request is a RDMA read or atomic, and the ACK is
 894                  * for a later operation, this ACK NAKs the RDMA read or
 895                  * atomic.  In other words, only a RDMA_READ_LAST or ONLY
 896                  * can ACK a RDMA read and likewise for atomic ops.  Note
 897                  * that the NAK case can only happen if relaxed ordering is
 898                  * used and requests are sent after an RDMA read or atomic
 899                  * is sent but before the response is received.
 900                  */
 901                 if ((wqe->wr.opcode == IB_WR_RDMA_READ &&
 902                      (opcode != OP(RDMA_READ_RESPONSE_LAST) || diff != 0)) ||
 903                     ((wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
 904                       wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) &&
 905                      (opcode != OP(ATOMIC_ACKNOWLEDGE) || diff != 0))) {
 906                         /*
 907                          * The last valid PSN seen is the previous
 908                          * request's.
 909                          */
 910                         update_last_psn(qp, wqe->psn - 1);
 911                         /* Retry this request. */
 912                         ipath_restart_rc(qp, wqe->psn, &wc);
 913                         /*
 914                          * No need to process the ACK/NAK since we are
 915                          * restarting an earlier request.
 916                          */
 917                         goto bail;
 918                 }
 919                 if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
 920                     wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)
 921                         *(u64 *) wqe->sg_list[0].vaddr = val;
 922                 if (qp->s_num_rd_atomic &&
 923                     (wqe->wr.opcode == IB_WR_RDMA_READ ||
 924                      wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
 925                      wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)) {
 926                         qp->s_num_rd_atomic--;
 927                         /* Restart sending task if fence is complete */
 928                         if ((qp->s_flags & IPATH_S_FENCE_PENDING) &&
 929                             !qp->s_num_rd_atomic) {
 930                                 qp->s_flags &= ~IPATH_S_FENCE_PENDING;
 931                                 tasklet_hi_schedule(&qp->s_task);
 932                         } else if (qp->s_flags & IPATH_S_RDMAR_PENDING) {
 933                                 qp->s_flags &= ~IPATH_S_RDMAR_PENDING;
 934                                 tasklet_hi_schedule(&qp->s_task);
 935                         }
 936                 }
 937                 /* Post a send completion queue entry if requested. */
 938                 if (!(qp->s_flags & IPATH_S_SIGNAL_REQ_WR) ||
 939                     (wqe->wr.send_flags & IB_SEND_SIGNALED)) {
 940                         wc.wr_id = wqe->wr.wr_id;
 941                         wc.status = IB_WC_SUCCESS;
 942                         wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
 943                         wc.vendor_err = 0;
 944                         wc.byte_len = wqe->length;
 945                         wc.imm_data = 0;
 946                         wc.qp = &qp->ibqp;
 947                         wc.src_qp = qp->remote_qpn;
 948                         wc.wc_flags = 0;
 949                         wc.pkey_index = 0;
 950                         wc.slid = qp->remote_ah_attr.dlid;
 951                         wc.sl = qp->remote_ah_attr.sl;
 952                         wc.dlid_path_bits = 0;
 953                         wc.port_num = 0;
 954                         ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 0);
 955                 }
 956                 qp->s_retry = qp->s_retry_cnt;
 957                 /*
 958                  * If we are completing a request which is in the process of
 959                  * being resent, we can stop resending it since we know the
 960                  * responder has already seen it.
 961                  */
 962                 if (qp->s_last == qp->s_cur) {
 963                         if (++qp->s_cur >= qp->s_size)
 964                                 qp->s_cur = 0;
 965                         qp->s_last = qp->s_cur;
 966                         if (qp->s_last == qp->s_tail)
 967                                 break;
 968                         wqe = get_swqe_ptr(qp, qp->s_cur);
 969                         qp->s_state = OP(SEND_LAST);
 970                         qp->s_psn = wqe->psn;
 971                 } else {
 972                         if (++qp->s_last >= qp->s_size)
 973                                 qp->s_last = 0;
 974                         if (qp->s_last == qp->s_tail)
 975                                 break;
 976                         wqe = get_swqe_ptr(qp, qp->s_last);
 977                 }
 978         }
 979
 980         switch (aeth >> 29) {
 981         case 0:         /* ACK */
 982                 dev->n_rc_acks++;
 983                 /* If this is a partial ACK, reset the retransmit timer. */
 984                 if (qp->s_last != qp->s_tail) {
 985                         spin_lock(&dev->pending_lock);
 986                         if (list_empty(&qp->timerwait))
 987                                 list_add_tail(&qp->timerwait,
 988                                         &dev->pending[dev->pending_index]);
 989                         spin_unlock(&dev->pending_lock);
 990                         /*
 991                          * If we get a partial ACK for a resent operation,
 992                          * we can stop resending the earlier packets and
 993                          * continue with the next packet the receiver wants.
 994                          */
 995                         if (ipath_cmp24(qp->s_psn, psn) <= 0) {
 996                                 reset_psn(qp, psn + 1);
 997                                 tasklet_hi_schedule(&qp->s_task);
 998                         }
 999                 } else if (ipath_cmp24(qp->s_psn, psn) <= 0) {
1000                         qp->s_state = OP(SEND_LAST);
1001                         qp->s_psn = psn + 1;
1002                 }
1003                 ipath_get_credit(qp, aeth);
1004                 qp->s_rnr_retry = qp->s_rnr_retry_cnt;
1005                 qp->s_retry = qp->s_retry_cnt;
1006                 update_last_psn(qp, psn);
1007                 ret = 1;
1008                 goto bail;
1009
1010         case 1:         /* RNR NAK */
1011                 dev->n_rnr_naks++;
1012                 if (qp->s_last == qp->s_tail)
1013                         goto bail;
1014                 if (qp->s_rnr_retry == 0) {
1015                         wc.status = IB_WC_RNR_RETRY_EXC_ERR;
1016                         goto class_b;
1017                 }
1018                 if (qp->s_rnr_retry_cnt < 7)
1019                         qp->s_rnr_retry--;
1020
1021                 /* The last valid PSN is the previous PSN. */
1022                 update_last_psn(qp, psn - 1);
1023
1024                 if (wqe->wr.opcode == IB_WR_RDMA_READ)
1025                         dev->n_rc_resends++;
1026                 else
1027                         dev->n_rc_resends +=
1028                                 (qp->s_psn - psn) & IPATH_PSN_MASK;
1029
1030                 reset_psn(qp, psn);
1031
1032                 qp->s_rnr_timeout =
1033                         ib_ipath_rnr_table[(aeth >> IPATH_AETH_CREDIT_SHIFT) &
1034                                            IPATH_AETH_CREDIT_MASK];
1035                 ipath_insert_rnr_queue(qp);
1036                 goto bail;
1037
1038         case 3:         /* NAK */
1039                 if (qp->s_last == qp->s_tail)
1040                         goto bail;
1041                 /* The last valid PSN is the previous PSN. */
1042                 update_last_psn(qp, psn - 1);
1043                 switch ((aeth >> IPATH_AETH_CREDIT_SHIFT) &
1044                         IPATH_AETH_CREDIT_MASK) {
1045                 case 0: /* PSN sequence error */
1046                         dev->n_seq_naks++;
1047                         /*
1048                          * Back up to the responder's expected PSN.
1049                          * Note that we might get a NAK in the middle of an
1050                          * RDMA READ response which terminates the RDMA
1051                          * READ.
1052                          */
1053                         ipath_restart_rc(qp, psn, &wc);
1054                         break;
1055
1056                 case 1: /* Invalid Request */
1057                         wc.status = IB_WC_REM_INV_REQ_ERR;
1058                         dev->n_other_naks++;
1059                         goto class_b;
1060
1061                 case 2: /* Remote Access Error */
1062                         wc.status = IB_WC_REM_ACCESS_ERR;
1063                         dev->n_other_naks++;
1064                         goto class_b;
1065
1066                 case 3: /* Remote Operation Error */
1067                         wc.status = IB_WC_REM_OP_ERR;
1068                         dev->n_other_naks++;
1069                 class_b:
1070                         wc.wr_id = wqe->wr.wr_id;
1071                         wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
1072                         wc.vendor_err = 0;
1073                         wc.byte_len = 0;
1074                         wc.qp = &qp->ibqp;
1075                         wc.imm_data = 0;
1076                         wc.src_qp = qp->remote_qpn;
1077                         wc.wc_flags = 0;
1078                         wc.pkey_index = 0;
1079                         wc.slid = qp->remote_ah_attr.dlid;
1080                         wc.sl = qp->remote_ah_attr.sl;
1081                         wc.dlid_path_bits = 0;
1082                         wc.port_num = 0;
1083                         ipath_sqerror_qp(qp, &wc);
1084                         break;
1085
1086                 default:
1087                         /* Ignore other reserved NAK error codes */
1088                         goto reserved;
1089                 }
1090                 qp->s_rnr_retry = qp->s_rnr_retry_cnt;
1091                 goto bail;
1092
1093         default:                /* 2: reserved */
1094         reserved:
1095                 /* Ignore reserved NAK codes. */
1096                 goto bail;
1097         }
1098
1099 bail:
1100         return ret;
1101 }
1102
1103 /**
1104  * ipath_rc_rcv_resp - process an incoming RC response packet
1105  * @dev: the device this packet came in on
1106  * @ohdr: the other headers for this packet
1107  * @data: the packet data
1108  * @tlen: the packet length
1109  * @qp: the QP for this packet
1110  * @opcode: the opcode for this packet
1111  * @psn: the packet sequence number for this packet
1112  * @hdrsize: the header length
1113  * @pmtu: the path MTU
1114  * @header_in_data: true if part of the header data is in the data buffer
1115  *
1116  * This is called from ipath_rc_rcv() to process an incoming RC response
1117  * packet for the given QP.
1118  * Called at interrupt level.
1119  */
1120 static inline void ipath_rc_rcv_resp(struct ipath_ibdev *dev,
1121                                      struct ipath_other_headers *ohdr,
1122                                      void *data, u32 tlen,
1123                                      struct ipath_qp *qp,
1124                                      u32 opcode,
1125                                      u32 psn, u32 hdrsize, u32 pmtu,
1126                                      int header_in_data)
1127 {
1128         struct ipath_swqe *wqe;
1129         unsigned long flags;
1130         struct ib_wc wc;
1131         int diff;
1132         u32 pad;
1133         u32 aeth;
1134         u64 val;
1135
1136         spin_lock_irqsave(&qp->s_lock, flags);
1137
1138         /* Ignore invalid responses. */
1139         if (ipath_cmp24(psn, qp->s_next_psn) >= 0)
1140                 goto ack_done;
1141
1142         /* Ignore duplicate responses. */
1143         diff = ipath_cmp24(psn, qp->s_last_psn);
1144         if (unlikely(diff <= 0)) {
1145                 /* Update credits for "ghost" ACKs */
1146                 if (diff == 0 && opcode == OP(ACKNOWLEDGE)) {
1147                         if (!header_in_data)
1148                                 aeth = be32_to_cpu(ohdr->u.aeth);
1149                         else {
1150                                 aeth = be32_to_cpu(((__be32 *) data)[0]);
1151                                 data += sizeof(__be32);
1152                         }
1153                         if ((aeth >> 29) == 0)
1154                                 ipath_get_credit(qp, aeth);
1155                 }
1156                 goto ack_done;
1157         }
1158
1159         if (unlikely(qp->s_last == qp->s_tail))
1160                 goto ack_done;
1161         wqe = get_swqe_ptr(qp, qp->s_last);
1162
1163         switch (opcode) {
1164         case OP(ACKNOWLEDGE):
1165         case OP(ATOMIC_ACKNOWLEDGE):
1166         case OP(RDMA_READ_RESPONSE_FIRST):
1167                 if (!header_in_data)
1168                         aeth = be32_to_cpu(ohdr->u.aeth);
1169                 else {
1170                         aeth = be32_to_cpu(((__be32 *) data)[0]);
1171                         data += sizeof(__be32);
1172                 }
1173                 if (opcode == OP(ATOMIC_ACKNOWLEDGE)) {
1174                         if (!header_in_data) {
1175                                 __be32 *p = ohdr->u.at.atomic_ack_eth;
1176
1177                                 val = ((u64) be32_to_cpu(p[0]) << 32) |
1178                                         be32_to_cpu(p[1]);
1179                         } else
1180                                 val = be64_to_cpu(((__be64 *) data)[0]);
1181                 } else
1182                         val = 0;
1183                 if (!do_rc_ack(qp, aeth, psn, opcode, val) ||
1184                     opcode != OP(RDMA_READ_RESPONSE_FIRST))
1185                         goto ack_done;
1186                 hdrsize += 4;
1187                 wqe = get_swqe_ptr(qp, qp->s_last);
1188                 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1189                         goto ack_op_err;
1190                 /*
1191                  * If this is a response to a resent RDMA read, we
1192                  * have to be careful to copy the data to the right
1193                  * location.
1194                  */
1195                 qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
1196                                                   wqe, psn, pmtu);
1197                 goto read_middle;
1198
1199         case OP(RDMA_READ_RESPONSE_MIDDLE):
1200                 /* no AETH, no ACK */
1201                 if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) {
1202                         dev->n_rdma_seq++;
1203                         ipath_restart_rc(qp, qp->s_last_psn + 1, &wc);
1204                         goto ack_done;
1205                 }
1206                 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1207                         goto ack_op_err;
1208         read_middle:
1209                 if (unlikely(tlen != (hdrsize + pmtu + 4)))
1210                         goto ack_len_err;
1211                 if (unlikely(pmtu >= qp->s_rdma_read_len))
1212                         goto ack_len_err;
1213
1214                 /* We got a response so update the timeout. */
1215                 spin_lock(&dev->pending_lock);
1216                 if (qp->s_rnr_timeout == 0 && !list_empty(&qp->timerwait))
1217                         list_move_tail(&qp->timerwait,
1218                                        &dev->pending[dev->pending_index]);
1219                 spin_unlock(&dev->pending_lock);
1220
1221                 if (opcode == OP(RDMA_READ_RESPONSE_MIDDLE))
1222                         qp->s_retry = qp->s_retry_cnt;
1223
1224                 /*
1225                  * Update the RDMA receive state but do the copy w/o
1226                  * holding the locks and blocking interrupts.
1227                  */
1228                 qp->s_rdma_read_len -= pmtu;
1229                 update_last_psn(qp, psn);
1230                 spin_unlock_irqrestore(&qp->s_lock, flags);
1231                 ipath_copy_sge(&qp->s_rdma_read_sge, data, pmtu);
1232                 goto bail;
1233
1234         case OP(RDMA_READ_RESPONSE_ONLY):
1235                 if (!header_in_data)
1236                         aeth = be32_to_cpu(ohdr->u.aeth);
1237                 else
1238                         aeth = be32_to_cpu(((__be32 *) data)[0]);
1239                 if (!do_rc_ack(qp, aeth, psn, opcode, 0))
1240                         goto ack_done;
1241                 /* Get the number of bytes the message was padded by. */
1242                 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1243                 /*
1244                  * Check that the data size is >= 0 && <= pmtu.
1245                  * Remember to account for the AETH header (4) and
1246                  * ICRC (4).
1247                  */
1248                 if (unlikely(tlen < (hdrsize + pad + 8)))
1249                         goto ack_len_err;
1250                 /*
1251                  * If this is a response to a resent RDMA read, we
1252                  * have to be careful to copy the data to the right
1253                  * location.
1254                  */
1255                 wqe = get_swqe_ptr(qp, qp->s_last);
1256                 qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
1257                                                   wqe, psn, pmtu);
1258                 goto read_last;
1259
1260         case OP(RDMA_READ_RESPONSE_LAST):
1261                 /* ACKs READ req. */
1262                 if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) {
1263                         dev->n_rdma_seq++;
1264                         ipath_restart_rc(qp, qp->s_last_psn + 1, &wc);
1265                         goto ack_done;
1266                 }
1267                 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1268                         goto ack_op_err;
1269                 /* Get the number of bytes the message was padded by. */
1270                 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1271                 /*
1272                  * Check that the data size is >= 1 && <= pmtu.
1273                  * Remember to account for the AETH header (4) and
1274                  * ICRC (4).
1275                  */
1276                 if (unlikely(tlen <= (hdrsize + pad + 8)))
1277                         goto ack_len_err;
1278         read_last:
1279                 tlen -= hdrsize + pad + 8;
1280                 if (unlikely(tlen != qp->s_rdma_read_len))
1281                         goto ack_len_err;
1282                 if (!header_in_data)
1283                         aeth = be32_to_cpu(ohdr->u.aeth);
1284                 else {
1285                         aeth = be32_to_cpu(((__be32 *) data)[0]);
1286                         data += sizeof(__be32);
1287                 }
1288                 ipath_copy_sge(&qp->s_rdma_read_sge, data, tlen);
1289                 (void) do_rc_ack(qp, aeth, psn,
1290                                  OP(RDMA_READ_RESPONSE_LAST), 0);
1291                 goto ack_done;
1292         }
1293
1294 ack_done:
1295         spin_unlock_irqrestore(&qp->s_lock, flags);
1296         goto bail;
1297
1298 ack_op_err:
1299         wc.status = IB_WC_LOC_QP_OP_ERR;
1300         goto ack_err;
1301
1302 ack_len_err:
1303         wc.status = IB_WC_LOC_LEN_ERR;
1304 ack_err:
1305         wc.wr_id = wqe->wr.wr_id;
1306         wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
1307         wc.vendor_err = 0;
1308         wc.byte_len = 0;
1309         wc.imm_data = 0;
1310         wc.qp = &qp->ibqp;
1311         wc.src_qp = qp->remote_qpn;
1312         wc.wc_flags = 0;
1313         wc.pkey_index = 0;
1314         wc.slid = qp->remote_ah_attr.dlid;
1315         wc.sl = qp->remote_ah_attr.sl;
1316         wc.dlid_path_bits = 0;
1317         wc.port_num = 0;
1318         ipath_sqerror_qp(qp, &wc);
1319         spin_unlock_irqrestore(&qp->s_lock, flags);
1320 bail:
1321         return;
1322 }
1323
1324 /**
1325  * ipath_rc_rcv_error - process an incoming duplicate or error RC packet
1326  * @dev: the device this packet came in on
1327  * @ohdr: the other headers for this packet
1328  * @data: the packet data
1329  * @qp: the QP for this packet
1330  * @opcode: the opcode for this packet
1331  * @psn: the packet sequence number for this packet
1332  * @diff: the difference between the PSN and the expected PSN
1333  * @header_in_data: true if part of the header data is in the data buffer
1334  *
1335  * This is called from ipath_rc_rcv() to process an unexpected
1336  * incoming RC packet for the given QP.
1337  * Called at interrupt level.
1338  * Return 1 if no more processing is needed; otherwise return 0 to
1339  * schedule a response to be sent.
1340  */
1341 static inline int ipath_rc_rcv_error(struct ipath_ibdev *dev,
1342                                      struct ipath_other_headers *ohdr,
1343                                      void *data,
1344                                      struct ipath_qp *qp,
1345                                      u32 opcode,
1346                                      u32 psn,
1347                                      int diff,
1348                                      int header_in_data)
1349 {
1350         struct ipath_ack_entry *e;
1351         u8 i, prev;
1352         int old_req;
1353         unsigned long flags;
1354
1355         if (diff > 0) {
1356                 /*
1357                  * Packet sequence error.
1358                  * A NAK will ACK earlier sends and RDMA writes.
1359                  * Don't queue the NAK if we already sent one.
1360                  */
1361                 if (!qp->r_nak_state) {
1362                         qp->r_nak_state = IB_NAK_PSN_ERROR;
1363                         /* Use the expected PSN. */
1364                         qp->r_ack_psn = qp->r_psn;
1365                         goto send_ack;
1366                 }
1367                 goto done;
1368         }
1369
1370         /*
1371          * Handle a duplicate request.  Don't re-execute SEND, RDMA
1372          * write or atomic op.  Don't NAK errors, just silently drop
1373          * the duplicate request.  Note that r_sge, r_len, and
1374          * r_rcv_len may be in use so don't modify them.
1375          *
1376          * We are supposed to ACK the earliest duplicate PSN but we
1377          * can coalesce an outstanding duplicate ACK.  We have to
1378          * send the earliest so that RDMA reads can be restarted at
1379          * the requester's expected PSN.
1380          *
1381          * First, find where this duplicate PSN falls within the
1382          * ACKs previously sent.
1383          */
1384         psn &= IPATH_PSN_MASK;
1385         e = NULL;
1386         old_req = 1;
1387         spin_lock_irqsave(&qp->s_lock, flags);
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         tasklet_hi_schedule(&qp->s_task);
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 static 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         qp->state = IB_QPS_ERR;
1533         lastwqe = ipath_error_qp(qp, err);
1534         spin_unlock_irqrestore(&qp->s_lock, flags);
1535
1536         if (lastwqe) {
1537                 struct ib_event ev;
1538
1539                 ev.device = qp->ibqp.device;
1540                 ev.element.qp = &qp->ibqp;
1541                 ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
1542                 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
1543         }
1544 }
1545
1546 static inline void ipath_update_ack_queue(struct ipath_qp *qp, unsigned n)
1547 {
1548         unsigned long flags;
1549         unsigned next;
1550
1551         next = n + 1;
1552         if (next > IPATH_MAX_RDMA_ATOMIC)
1553                 next = 0;
1554         spin_lock_irqsave(&qp->s_lock, flags);
1555         if (n == qp->s_tail_ack_queue) {
1556                 qp->s_tail_ack_queue = next;
1557                 qp->s_ack_state = OP(ACKNOWLEDGE);
1558         }
1559         spin_unlock_irqrestore(&qp->s_lock, flags);
1560 }
1561
1562 /**
1563  * ipath_rc_rcv - process an incoming RC packet
1564  * @dev: the device this packet came in on
1565  * @hdr: the header of this packet
1566  * @has_grh: true if the header has a GRH
1567  * @data: the packet data
1568  * @tlen: the packet length
1569  * @qp: the QP for this packet
1570  *
1571  * This is called from ipath_qp_rcv() to process an incoming RC packet
1572  * for the given QP.
1573  * Called at interrupt level.
1574  */
1575 void ipath_rc_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
1576                   int has_grh, void *data, u32 tlen, struct ipath_qp *qp)
1577 {
1578         struct ipath_other_headers *ohdr;
1579         u32 opcode;
1580         u32 hdrsize;
1581         u32 psn;
1582         u32 pad;
1583         struct ib_wc wc;
1584         u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
1585         int diff;
1586         struct ib_reth *reth;
1587         int header_in_data;
1588
1589         /* Validate the SLID. See Ch. 9.6.1.5 */
1590         if (unlikely(be16_to_cpu(hdr->lrh[3]) != qp->remote_ah_attr.dlid))
1591                 goto done;
1592
1593         /* Check for GRH */
1594         if (!has_grh) {
1595                 ohdr = &hdr->u.oth;
1596                 hdrsize = 8 + 12;       /* LRH + BTH */
1597                 psn = be32_to_cpu(ohdr->bth[2]);
1598                 header_in_data = 0;
1599         } else {
1600                 ohdr = &hdr->u.l.oth;
1601                 hdrsize = 8 + 40 + 12;  /* LRH + GRH + BTH */
1602                 /*
1603                  * The header with GRH is 60 bytes and the core driver sets
1604                  * the eager header buffer size to 56 bytes so the last 4
1605                  * bytes of the BTH header (PSN) is in the data buffer.
1606                  */
1607                 header_in_data = dev->dd->ipath_rcvhdrentsize == 16;
1608                 if (header_in_data) {
1609                         psn = be32_to_cpu(((__be32 *) data)[0]);
1610                         data += sizeof(__be32);
1611                 } else
1612                         psn = be32_to_cpu(ohdr->bth[2]);
1613         }
1614
1615         /*
1616          * Process responses (ACKs) before anything else.  Note that the
1617          * packet sequence number will be for something in the send work
1618          * queue rather than the expected receive packet sequence number.
1619          * In other words, this QP is the requester.
1620          */
1621         opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
1622         if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
1623             opcode <= OP(ATOMIC_ACKNOWLEDGE)) {
1624                 ipath_rc_rcv_resp(dev, ohdr, data, tlen, qp, opcode, psn,
1625                                   hdrsize, pmtu, header_in_data);
1626                 goto done;
1627         }
1628
1629         /* Compute 24 bits worth of difference. */
1630         diff = ipath_cmp24(psn, qp->r_psn);
1631         if (unlikely(diff)) {
1632                 if (ipath_rc_rcv_error(dev, ohdr, data, qp, opcode,
1633                                        psn, diff, header_in_data))
1634                         goto done;
1635                 goto send_ack;
1636         }
1637
1638         /* Check for opcode sequence errors. */
1639         switch (qp->r_state) {
1640         case OP(SEND_FIRST):
1641         case OP(SEND_MIDDLE):
1642                 if (opcode == OP(SEND_MIDDLE) ||
1643                     opcode == OP(SEND_LAST) ||
1644                     opcode == OP(SEND_LAST_WITH_IMMEDIATE))
1645                         break;
1646         nack_inv:
1647                 ipath_rc_error(qp, IB_WC_REM_INV_REQ_ERR);
1648                 qp->r_nak_state = IB_NAK_INVALID_REQUEST;
1649                 qp->r_ack_psn = qp->r_psn;
1650                 goto send_ack;
1651
1652         case OP(RDMA_WRITE_FIRST):
1653         case OP(RDMA_WRITE_MIDDLE):
1654                 if (opcode == OP(RDMA_WRITE_MIDDLE) ||
1655                     opcode == OP(RDMA_WRITE_LAST) ||
1656                     opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
1657                         break;
1658                 goto nack_inv;
1659
1660         default:
1661                 if (opcode == OP(SEND_MIDDLE) ||
1662                     opcode == OP(SEND_LAST) ||
1663                     opcode == OP(SEND_LAST_WITH_IMMEDIATE) ||
1664                     opcode == OP(RDMA_WRITE_MIDDLE) ||
1665                     opcode == OP(RDMA_WRITE_LAST) ||
1666                     opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
1667                         goto nack_inv;
1668                 /*
1669                  * Note that it is up to the requester to not send a new
1670                  * RDMA read or atomic operation before receiving an ACK
1671                  * for the previous operation.
1672                  */
1673                 break;
1674         }
1675
1676         wc.imm_data = 0;
1677         wc.wc_flags = 0;
1678
1679         /* OK, process the packet. */
1680         switch (opcode) {
1681         case OP(SEND_FIRST):
1682                 if (!ipath_get_rwqe(qp, 0)) {
1683                 rnr_nak:
1684                         qp->r_nak_state = IB_RNR_NAK | qp->r_min_rnr_timer;
1685                         qp->r_ack_psn = qp->r_psn;
1686                         goto send_ack;
1687                 }
1688                 qp->r_rcv_len = 0;
1689                 /* FALLTHROUGH */
1690         case OP(SEND_MIDDLE):
1691         case OP(RDMA_WRITE_MIDDLE):
1692         send_middle:
1693                 /* Check for invalid length PMTU or posted rwqe len. */
1694                 if (unlikely(tlen != (hdrsize + pmtu + 4)))
1695                         goto nack_inv;
1696                 qp->r_rcv_len += pmtu;
1697                 if (unlikely(qp->r_rcv_len > qp->r_len))
1698                         goto nack_inv;
1699                 ipath_copy_sge(&qp->r_sge, data, pmtu);
1700                 break;
1701
1702         case OP(RDMA_WRITE_LAST_WITH_IMMEDIATE):
1703                 /* consume RWQE */
1704                 if (!ipath_get_rwqe(qp, 1))
1705                         goto rnr_nak;
1706                 goto send_last_imm;
1707
1708         case OP(SEND_ONLY):
1709         case OP(SEND_ONLY_WITH_IMMEDIATE):
1710                 if (!ipath_get_rwqe(qp, 0))
1711                         goto rnr_nak;
1712                 qp->r_rcv_len = 0;
1713                 if (opcode == OP(SEND_ONLY))
1714                         goto send_last;
1715                 /* FALLTHROUGH */
1716         case OP(SEND_LAST_WITH_IMMEDIATE):
1717         send_last_imm:
1718                 if (header_in_data) {
1719                         wc.imm_data = *(__be32 *) data;
1720                         data += sizeof(__be32);
1721                 } else {
1722                         /* Immediate data comes after BTH */
1723                         wc.imm_data = ohdr->u.imm_data;
1724                 }
1725                 hdrsize += 4;
1726                 wc.wc_flags = IB_WC_WITH_IMM;
1727                 /* FALLTHROUGH */
1728         case OP(SEND_LAST):
1729         case OP(RDMA_WRITE_LAST):
1730         send_last:
1731                 /* Get the number of bytes the message was padded by. */
1732                 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1733                 /* Check for invalid length. */
1734                 /* XXX LAST len should be >= 1 */
1735                 if (unlikely(tlen < (hdrsize + pad + 4)))
1736                         goto nack_inv;
1737                 /* Don't count the CRC. */
1738                 tlen -= (hdrsize + pad + 4);
1739                 wc.byte_len = tlen + qp->r_rcv_len;
1740                 if (unlikely(wc.byte_len > qp->r_len))
1741                         goto nack_inv;
1742                 ipath_copy_sge(&qp->r_sge, data, tlen);
1743                 qp->r_msn++;
1744                 if (!qp->r_wrid_valid)
1745                         break;
1746                 qp->r_wrid_valid = 0;
1747                 wc.wr_id = qp->r_wr_id;
1748                 wc.status = IB_WC_SUCCESS;
1749                 wc.opcode = IB_WC_RECV;
1750                 wc.vendor_err = 0;
1751                 wc.qp = &qp->ibqp;
1752                 wc.src_qp = qp->remote_qpn;
1753                 wc.pkey_index = 0;
1754                 wc.slid = qp->remote_ah_attr.dlid;
1755                 wc.sl = qp->remote_ah_attr.sl;
1756                 wc.dlid_path_bits = 0;
1757                 wc.port_num = 0;
1758                 /* Signal completion event if the solicited bit is set. */
1759                 ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
1760                                (ohdr->bth[0] &
1761                                 __constant_cpu_to_be32(1 << 23)) != 0);
1762                 break;
1763
1764         case OP(RDMA_WRITE_FIRST):
1765         case OP(RDMA_WRITE_ONLY):
1766         case OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE):
1767                 if (unlikely(!(qp->qp_access_flags &
1768                                IB_ACCESS_REMOTE_WRITE)))
1769                         goto nack_inv;
1770                 /* consume RWQE */
1771                 /* RETH comes after BTH */
1772                 if (!header_in_data)
1773                         reth = &ohdr->u.rc.reth;
1774                 else {
1775                         reth = (struct ib_reth *)data;
1776                         data += sizeof(*reth);
1777                 }
1778                 hdrsize += sizeof(*reth);
1779                 qp->r_len = be32_to_cpu(reth->length);
1780                 qp->r_rcv_len = 0;
1781                 if (qp->r_len != 0) {
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, &qp->r_sge,
1788                                            qp->r_len, vaddr, rkey,
1789                                            IB_ACCESS_REMOTE_WRITE);
1790                         if (unlikely(!ok))
1791                                 goto nack_acc;
1792                 } else {
1793                         qp->r_sge.sg_list = NULL;
1794                         qp->r_sge.sge.mr = NULL;
1795                         qp->r_sge.sge.vaddr = NULL;
1796                         qp->r_sge.sge.length = 0;
1797                         qp->r_sge.sge.sge_length = 0;
1798                 }
1799                 if (opcode == OP(RDMA_WRITE_FIRST))
1800                         goto send_middle;
1801                 else if (opcode == OP(RDMA_WRITE_ONLY))
1802                         goto send_last;
1803                 if (!ipath_get_rwqe(qp, 1))
1804                         goto rnr_nak;
1805                 goto send_last_imm;
1806
1807         case OP(RDMA_READ_REQUEST): {
1808                 struct ipath_ack_entry *e;
1809                 u32 len;
1810                 u8 next;
1811
1812                 if (unlikely(!(qp->qp_access_flags &
1813                                IB_ACCESS_REMOTE_READ)))
1814                         goto nack_inv;
1815                 next = qp->r_head_ack_queue + 1;
1816                 if (next > IPATH_MAX_RDMA_ATOMIC)
1817                         next = 0;
1818                 if (unlikely(next == qp->s_tail_ack_queue)) {
1819                         if (!qp->s_ack_queue[next].sent)
1820                                 goto nack_inv;
1821                         ipath_update_ack_queue(qp, next);
1822                 }
1823                 e = &qp->s_ack_queue[qp->r_head_ack_queue];
1824                 /* RETH comes after BTH */
1825                 if (!header_in_data)
1826                         reth = &ohdr->u.rc.reth;
1827                 else {
1828                         reth = (struct ib_reth *)data;
1829                         data += sizeof(*reth);
1830                 }
1831                 len = be32_to_cpu(reth->length);
1832                 if (len) {
1833                         u32 rkey = be32_to_cpu(reth->rkey);
1834                         u64 vaddr = be64_to_cpu(reth->vaddr);
1835                         int ok;
1836
1837                         /* Check rkey & NAK */
1838                         ok = ipath_rkey_ok(qp, &e->rdma_sge, len, vaddr,
1839                                            rkey, IB_ACCESS_REMOTE_READ);
1840                         if (unlikely(!ok))
1841                                 goto nack_acc;
1842                         /*
1843                          * Update the next expected PSN.  We add 1 later
1844                          * below, so only add the remainder here.
1845                          */
1846                         if (len > pmtu)
1847                                 qp->r_psn += (len - 1) / pmtu;
1848                 } else {
1849                         e->rdma_sge.sg_list = NULL;
1850                         e->rdma_sge.num_sge = 0;
1851                         e->rdma_sge.sge.mr = NULL;
1852                         e->rdma_sge.sge.vaddr = NULL;
1853                         e->rdma_sge.sge.length = 0;
1854                         e->rdma_sge.sge.sge_length = 0;
1855                 }
1856                 e->opcode = opcode;
1857                 e->sent = 0;
1858                 e->psn = psn;
1859                 /*
1860                  * We need to increment the MSN here instead of when we
1861                  * finish sending the result since a duplicate request would
1862                  * increment it more than once.
1863                  */
1864                 qp->r_msn++;
1865                 qp->r_psn++;
1866                 qp->r_state = opcode;
1867                 qp->r_nak_state = 0;
1868                 barrier();
1869                 qp->r_head_ack_queue = next;
1870
1871                 /* Call ipath_do_rc_send() in another thread. */
1872                 tasklet_hi_schedule(&qp->s_task);
1873
1874                 goto done;
1875         }
1876
1877         case OP(COMPARE_SWAP):
1878         case OP(FETCH_ADD): {
1879                 struct ib_atomic_eth *ateth;
1880                 struct ipath_ack_entry *e;
1881                 u64 vaddr;
1882                 atomic64_t *maddr;
1883                 u64 sdata;
1884                 u32 rkey;
1885                 u8 next;
1886
1887                 if (unlikely(!(qp->qp_access_flags &
1888                                IB_ACCESS_REMOTE_ATOMIC)))
1889                         goto nack_inv;
1890                 next = qp->r_head_ack_queue + 1;
1891                 if (next > IPATH_MAX_RDMA_ATOMIC)
1892                         next = 0;
1893                 if (unlikely(next == qp->s_tail_ack_queue)) {
1894                         if (!qp->s_ack_queue[next].sent)
1895                                 goto nack_inv;
1896                         ipath_update_ack_queue(qp, next);
1897                 }
1898                 if (!header_in_data)
1899                         ateth = &ohdr->u.atomic_eth;
1900                 else
1901                         ateth = (struct ib_atomic_eth *)data;
1902                 vaddr = ((u64) be32_to_cpu(ateth->vaddr[0]) << 32) |
1903                         be32_to_cpu(ateth->vaddr[1]);
1904                 if (unlikely(vaddr & (sizeof(u64) - 1)))
1905                         goto nack_inv;
1906                 rkey = be32_to_cpu(ateth->rkey);
1907                 /* Check rkey & NAK */
1908                 if (unlikely(!ipath_rkey_ok(qp, &qp->r_sge,
1909                                             sizeof(u64), vaddr, rkey,
1910                                             IB_ACCESS_REMOTE_ATOMIC)))
1911                         goto nack_acc;
1912                 /* Perform atomic OP and save result. */
1913                 maddr = (atomic64_t *) qp->r_sge.sge.vaddr;
1914                 sdata = be64_to_cpu(ateth->swap_data);
1915                 e = &qp->s_ack_queue[qp->r_head_ack_queue];
1916                 e->atomic_data = (opcode == OP(FETCH_ADD)) ?
1917                         (u64) atomic64_add_return(sdata, maddr) - sdata :
1918                         (u64) cmpxchg((u64 *) qp->r_sge.sge.vaddr,
1919                                       be64_to_cpu(ateth->compare_data),
1920                                       sdata);
1921                 e->opcode = opcode;
1922                 e->sent = 0;
1923                 e->psn = psn & IPATH_PSN_MASK;
1924                 qp->r_msn++;
1925                 qp->r_psn++;
1926                 qp->r_state = opcode;
1927                 qp->r_nak_state = 0;
1928                 barrier();
1929                 qp->r_head_ack_queue = next;
1930
1931                 /* Call ipath_do_rc_send() in another thread. */
1932                 tasklet_hi_schedule(&qp->s_task);
1933
1934                 goto done;
1935         }
1936
1937         default:
1938                 /* NAK unknown opcodes. */
1939                 goto nack_inv;
1940         }
1941         qp->r_psn++;
1942         qp->r_state = opcode;
1943         qp->r_ack_psn = psn;
1944         qp->r_nak_state = 0;
1945         /* Send an ACK if requested or required. */
1946         if (psn & (1 << 31))
1947                 goto send_ack;
1948         goto done;
1949
1950 nack_acc:
1951         ipath_rc_error(qp, IB_WC_REM_ACCESS_ERR);
1952         qp->r_nak_state = IB_NAK_REMOTE_ACCESS_ERROR;
1953         qp->r_ack_psn = qp->r_psn;
1954
1955 send_ack:
1956         send_rc_ack(qp);
1957
1958 done:
1959         return;
1960 }