drivers/infiniband/hw/qib/qib_ruc.c

   1 /*
   2  * Copyright (c) 2006, 2007, 2008, 2009 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/spinlock.h>
  35 #include <rdma/ib_smi.h>
  36
  37 #include "qib.h"
  38 #include "qib_mad.h"
  39
  40 /*
  41  * Switch to alternate path.
  42  * The QP s_lock should be held and interrupts disabled.
  43  */
  44 void qib_migrate_qp(struct rvt_qp *qp)
  45 {
  46         struct ib_event ev;
  47
  48         qp->s_mig_state = IB_MIG_MIGRATED;
  49         qp->remote_ah_attr = qp->alt_ah_attr;
  50         qp->port_num = rdma_ah_get_port_num(&qp->alt_ah_attr);
  51         qp->s_pkey_index = qp->s_alt_pkey_index;
  52
  53         ev.device = qp->ibqp.device;
  54         ev.element.qp = &qp->ibqp;
  55         ev.event = IB_EVENT_PATH_MIG;
  56         qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
  57 }
  58
  59 static __be64 get_sguid(struct qib_ibport *ibp, unsigned index)
  60 {
  61         if (!index) {
  62                 struct qib_pportdata *ppd = ppd_from_ibp(ibp);
  63
  64                 return ppd->guid;
  65         }
  66         return ibp->guids[index - 1];
  67 }
  68
  69 static int gid_ok(union ib_gid *gid, __be64 gid_prefix, __be64 id)
  70 {
  71         return (gid->global.interface_id == id &&
  72                 (gid->global.subnet_prefix == gid_prefix ||
  73                  gid->global.subnet_prefix == IB_DEFAULT_GID_PREFIX));
  74 }
  75
  76 /*
  77  *
  78  * This should be called with the QP r_lock held.
  79  *
  80  * The s_lock will be acquired around the qib_migrate_qp() call.
  81  */
  82 int qib_ruc_check_hdr(struct qib_ibport *ibp, struct ib_header *hdr,
  83                       int has_grh, struct rvt_qp *qp, u32 bth0)
  84 {
  85         __be64 guid;
  86         unsigned long flags;
  87
  88         if (qp->s_mig_state == IB_MIG_ARMED && (bth0 & IB_BTH_MIG_REQ)) {
  89                 if (!has_grh) {
  90                         if (rdma_ah_get_ah_flags(&qp->alt_ah_attr) &
  91                             IB_AH_GRH)
  92                                 goto err;
  93                 } else {
  94                         const struct ib_global_route *grh;
  95
  96                         if (!(rdma_ah_get_ah_flags(&qp->alt_ah_attr) &
  97                               IB_AH_GRH))
  98                                 goto err;
  99                         grh = rdma_ah_read_grh(&qp->alt_ah_attr);
 100                         guid = get_sguid(ibp, grh->sgid_index);
 101                         if (!gid_ok(&hdr->u.l.grh.dgid,
 102                                     ibp->rvp.gid_prefix, guid))
 103                                 goto err;
 104                         if (!gid_ok(&hdr->u.l.grh.sgid,
 105                             grh->dgid.global.subnet_prefix,
 106                             grh->dgid.global.interface_id))
 107                                 goto err;
 108                 }
 109                 if (!qib_pkey_ok((u16)bth0,
 110                                  qib_get_pkey(ibp, qp->s_alt_pkey_index))) {
 111                         qib_bad_pkey(ibp,
 112                                      (u16)bth0,
 113                                      (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF,
 114                                      0, qp->ibqp.qp_num,
 115                                      hdr->lrh[3], hdr->lrh[1]);
 116                         goto err;
 117                 }
 118                 /* Validate the SLID. See Ch. 9.6.1.5 and 17.2.8 */
 119                 if ((be16_to_cpu(hdr->lrh[3]) !=
 120                      rdma_ah_get_dlid(&qp->alt_ah_attr)) ||
 121                     ppd_from_ibp(ibp)->port !=
 122                             rdma_ah_get_port_num(&qp->alt_ah_attr))
 123                         goto err;
 124                 spin_lock_irqsave(&qp->s_lock, flags);
 125                 qib_migrate_qp(qp);
 126                 spin_unlock_irqrestore(&qp->s_lock, flags);
 127         } else {
 128                 if (!has_grh) {
 129                         if (rdma_ah_get_ah_flags(&qp->remote_ah_attr) &
 130                             IB_AH_GRH)
 131                                 goto err;
 132                 } else {
 133                         const struct ib_global_route *grh;
 134
 135                         if (!(rdma_ah_get_ah_flags(&qp->remote_ah_attr) &
 136                               IB_AH_GRH))
 137                                 goto err;
 138                         grh = rdma_ah_read_grh(&qp->remote_ah_attr);
 139                         guid = get_sguid(ibp, grh->sgid_index);
 140                         if (!gid_ok(&hdr->u.l.grh.dgid,
 141                                     ibp->rvp.gid_prefix, guid))
 142                                 goto err;
 143                         if (!gid_ok(&hdr->u.l.grh.sgid,
 144                             grh->dgid.global.subnet_prefix,
 145                             grh->dgid.global.interface_id))
 146                                 goto err;
 147                 }
 148                 if (!qib_pkey_ok((u16)bth0,
 149                                  qib_get_pkey(ibp, qp->s_pkey_index))) {
 150                         qib_bad_pkey(ibp,
 151                                      (u16)bth0,
 152                                      (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF,
 153                                      0, qp->ibqp.qp_num,
 154                                      hdr->lrh[3], hdr->lrh[1]);
 155                         goto err;
 156                 }
 157                 /* Validate the SLID. See Ch. 9.6.1.5 */
 158                 if (be16_to_cpu(hdr->lrh[3]) !=
 159                     rdma_ah_get_dlid(&qp->remote_ah_attr) ||
 160                     ppd_from_ibp(ibp)->port != qp->port_num)
 161                         goto err;
 162                 if (qp->s_mig_state == IB_MIG_REARM &&
 163                     !(bth0 & IB_BTH_MIG_REQ))
 164                         qp->s_mig_state = IB_MIG_ARMED;
 165         }
 166
 167         return 0;
 168
 169 err:
 170         return 1;
 171 }
 172
 173 /**
 174  * qib_ruc_loopback - handle UC and RC lookback requests
 175  * @sqp: the sending QP
 176  *
 177  * This is called from qib_do_send() to
 178  * forward a WQE addressed to the same HCA.
 179  * Note that although we are single threaded due to the tasklet, we still
 180  * have to protect against post_send().  We don't have to worry about
 181  * receive interrupts since this is a connected protocol and all packets
 182  * will pass through here.
 183  */
 184 static void qib_ruc_loopback(struct rvt_qp *sqp)
 185 {
 186         struct qib_ibport *ibp = to_iport(sqp->ibqp.device, sqp->port_num);
 187         struct qib_pportdata *ppd = ppd_from_ibp(ibp);
 188         struct qib_devdata *dd = ppd->dd;
 189         struct rvt_dev_info *rdi = &dd->verbs_dev.rdi;
 190         struct rvt_qp *qp;
 191         struct rvt_swqe *wqe;
 192         struct rvt_sge *sge;
 193         unsigned long flags;
 194         struct ib_wc wc;
 195         u64 sdata;
 196         atomic64_t *maddr;
 197         enum ib_wc_status send_status;
 198         int release;
 199         int ret;
 200
 201         rcu_read_lock();
 202         /*
 203          * Note that we check the responder QP state after
 204          * checking the requester's state.
 205          */
 206         qp = rvt_lookup_qpn(rdi, &ibp->rvp, sqp->remote_qpn);
 207         if (!qp)
 208                 goto done;
 209
 210         spin_lock_irqsave(&sqp->s_lock, flags);
 211
 212         /* Return if we are already busy processing a work request. */
 213         if ((sqp->s_flags & (RVT_S_BUSY | RVT_S_ANY_WAIT)) ||
 214             !(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_OR_FLUSH_SEND))
 215                 goto unlock;
 216
 217         sqp->s_flags |= RVT_S_BUSY;
 218
 219 again:
 220         if (sqp->s_last == READ_ONCE(sqp->s_head))
 221                 goto clr_busy;
 222         wqe = rvt_get_swqe_ptr(sqp, sqp->s_last);
 223
 224         /* Return if it is not OK to start a new work reqeust. */
 225         if (!(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_NEXT_SEND_OK)) {
 226                 if (!(ib_rvt_state_ops[sqp->state] & RVT_FLUSH_SEND))
 227                         goto clr_busy;
 228                 /* We are in the error state, flush the work request. */
 229                 send_status = IB_WC_WR_FLUSH_ERR;
 230                 goto flush_send;
 231         }
 232
 233         /*
 234          * We can rely on the entry not changing without the s_lock
 235          * being held until we update s_last.
 236          * We increment s_cur to indicate s_last is in progress.
 237          */
 238         if (sqp->s_last == sqp->s_cur) {
 239                 if (++sqp->s_cur >= sqp->s_size)
 240                         sqp->s_cur = 0;
 241         }
 242         spin_unlock_irqrestore(&sqp->s_lock, flags);
 243
 244         if (!qp || !(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) ||
 245             qp->ibqp.qp_type != sqp->ibqp.qp_type) {
 246                 ibp->rvp.n_pkt_drops++;
 247                 /*
 248                  * For RC, the requester would timeout and retry so
 249                  * shortcut the timeouts and just signal too many retries.
 250                  */
 251                 if (sqp->ibqp.qp_type == IB_QPT_RC)
 252                         send_status = IB_WC_RETRY_EXC_ERR;
 253                 else
 254                         send_status = IB_WC_SUCCESS;
 255                 goto serr;
 256         }
 257
 258         memset(&wc, 0, sizeof(wc));
 259         send_status = IB_WC_SUCCESS;
 260
 261         release = 1;
 262         sqp->s_sge.sge = wqe->sg_list[0];
 263         sqp->s_sge.sg_list = wqe->sg_list + 1;
 264         sqp->s_sge.num_sge = wqe->wr.num_sge;
 265         sqp->s_len = wqe->length;
 266         switch (wqe->wr.opcode) {
 267         case IB_WR_SEND_WITH_IMM:
 268                 wc.wc_flags = IB_WC_WITH_IMM;
 269                 wc.ex.imm_data = wqe->wr.ex.imm_data;
 270                 /* FALLTHROUGH */
 271         case IB_WR_SEND:
 272                 ret = rvt_get_rwqe(qp, false);
 273                 if (ret < 0)
 274                         goto op_err;
 275                 if (!ret)
 276                         goto rnr_nak;
 277                 if (wqe->length > qp->r_len)
 278                         goto inv_err;
 279                 break;
 280
 281         case IB_WR_RDMA_WRITE_WITH_IMM:
 282                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
 283                         goto inv_err;
 284                 wc.wc_flags = IB_WC_WITH_IMM;
 285                 wc.ex.imm_data = wqe->wr.ex.imm_data;
 286                 ret = rvt_get_rwqe(qp, true);
 287                 if (ret < 0)
 288                         goto op_err;
 289                 if (!ret)
 290                         goto rnr_nak;
 291                 /* FALLTHROUGH */
 292         case IB_WR_RDMA_WRITE:
 293                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
 294                         goto inv_err;
 295                 if (wqe->length == 0)
 296                         break;
 297                 if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, wqe->length,
 298                                           wqe->rdma_wr.remote_addr,
 299                                           wqe->rdma_wr.rkey,
 300                                           IB_ACCESS_REMOTE_WRITE)))
 301                         goto acc_err;
 302                 qp->r_sge.sg_list = NULL;
 303                 qp->r_sge.num_sge = 1;
 304                 qp->r_sge.total_len = wqe->length;
 305                 break;
 306
 307         case IB_WR_RDMA_READ:
 308                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ)))
 309                         goto inv_err;
 310                 if (unlikely(!rvt_rkey_ok(qp, &sqp->s_sge.sge, wqe->length,
 311                                           wqe->rdma_wr.remote_addr,
 312                                           wqe->rdma_wr.rkey,
 313                                           IB_ACCESS_REMOTE_READ)))
 314                         goto acc_err;
 315                 release = 0;
 316                 sqp->s_sge.sg_list = NULL;
 317                 sqp->s_sge.num_sge = 1;
 318                 qp->r_sge.sge = wqe->sg_list[0];
 319                 qp->r_sge.sg_list = wqe->sg_list + 1;
 320                 qp->r_sge.num_sge = wqe->wr.num_sge;
 321                 qp->r_sge.total_len = wqe->length;
 322                 break;
 323
 324         case IB_WR_ATOMIC_CMP_AND_SWP:
 325         case IB_WR_ATOMIC_FETCH_AND_ADD:
 326                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC)))
 327                         goto inv_err;
 328                 if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, sizeof(u64),
 329                                           wqe->atomic_wr.remote_addr,
 330                                           wqe->atomic_wr.rkey,
 331                                           IB_ACCESS_REMOTE_ATOMIC)))
 332                         goto acc_err;
 333                 /* Perform atomic OP and save result. */
 334                 maddr = (atomic64_t *) qp->r_sge.sge.vaddr;
 335                 sdata = wqe->atomic_wr.compare_add;
 336                 *(u64 *) sqp->s_sge.sge.vaddr =
 337                         (wqe->atomic_wr.wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) ?
 338                         (u64) atomic64_add_return(sdata, maddr) - sdata :
 339                         (u64) cmpxchg((u64 *) qp->r_sge.sge.vaddr,
 340                                       sdata, wqe->atomic_wr.swap);
 341                 rvt_put_mr(qp->r_sge.sge.mr);
 342                 qp->r_sge.num_sge = 0;
 343                 goto send_comp;
 344
 345         default:
 346                 send_status = IB_WC_LOC_QP_OP_ERR;
 347                 goto serr;
 348         }
 349
 350         sge = &sqp->s_sge.sge;
 351         while (sqp->s_len) {
 352                 u32 len = sqp->s_len;
 353
 354                 if (len > sge->length)
 355                         len = sge->length;
 356                 if (len > sge->sge_length)
 357                         len = sge->sge_length;
 358                 BUG_ON(len == 0);
 359                 qib_copy_sge(&qp->r_sge, sge->vaddr, len, release);
 360                 sge->vaddr += len;
 361                 sge->length -= len;
 362                 sge->sge_length -= len;
 363                 if (sge->sge_length == 0) {
 364                         if (!release)
 365                                 rvt_put_mr(sge->mr);
 366                         if (--sqp->s_sge.num_sge)
 367                                 *sge = *sqp->s_sge.sg_list++;
 368                 } else if (sge->length == 0 && sge->mr->lkey) {
 369                         if (++sge->n >= RVT_SEGSZ) {
 370                                 if (++sge->m >= sge->mr->mapsz)
 371                                         break;
 372                                 sge->n = 0;
 373                         }
 374                         sge->vaddr =
 375                                 sge->mr->map[sge->m]->segs[sge->n].vaddr;
 376                         sge->length =
 377                                 sge->mr->map[sge->m]->segs[sge->n].length;
 378                 }
 379                 sqp->s_len -= len;
 380         }
 381         if (release)
 382                 rvt_put_ss(&qp->r_sge);
 383
 384         if (!test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags))
 385                 goto send_comp;
 386
 387         if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM)
 388                 wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
 389         else
 390                 wc.opcode = IB_WC_RECV;
 391         wc.wr_id = qp->r_wr_id;
 392         wc.status = IB_WC_SUCCESS;
 393         wc.byte_len = wqe->length;
 394         wc.qp = &qp->ibqp;
 395         wc.src_qp = qp->remote_qpn;
 396         wc.slid = rdma_ah_get_dlid(&qp->remote_ah_attr);
 397         wc.sl = rdma_ah_get_sl(&qp->remote_ah_attr);
 398         wc.port_num = 1;
 399         /* Signal completion event if the solicited bit is set. */
 400         rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc,
 401                      wqe->wr.send_flags & IB_SEND_SOLICITED);
 402
 403 send_comp:
 404         spin_lock_irqsave(&sqp->s_lock, flags);
 405         ibp->rvp.n_loop_pkts++;
 406 flush_send:
 407         sqp->s_rnr_retry = sqp->s_rnr_retry_cnt;
 408         qib_send_complete(sqp, wqe, send_status);
 409         goto again;
 410
 411 rnr_nak:
 412         /* Handle RNR NAK */
 413         if (qp->ibqp.qp_type == IB_QPT_UC)
 414                 goto send_comp;
 415         ibp->rvp.n_rnr_naks++;
 416         /*
 417          * Note: we don't need the s_lock held since the BUSY flag
 418          * makes this single threaded.
 419          */
 420         if (sqp->s_rnr_retry == 0) {
 421                 send_status = IB_WC_RNR_RETRY_EXC_ERR;
 422                 goto serr;
 423         }
 424         if (sqp->s_rnr_retry_cnt < 7)
 425                 sqp->s_rnr_retry--;
 426         spin_lock_irqsave(&sqp->s_lock, flags);
 427         if (!(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_RECV_OK))
 428                 goto clr_busy;
 429         rvt_add_rnr_timer(sqp, qp->r_min_rnr_timer <<
 430                                 IB_AETH_CREDIT_SHIFT);
 431         goto clr_busy;
 432
 433 op_err:
 434         send_status = IB_WC_REM_OP_ERR;
 435         wc.status = IB_WC_LOC_QP_OP_ERR;
 436         goto err;
 437
 438 inv_err:
 439         send_status =
 440                 sqp->ibqp.qp_type == IB_QPT_RC ?
 441                         IB_WC_REM_INV_REQ_ERR :
 442                         IB_WC_SUCCESS;
 443         wc.status = IB_WC_LOC_QP_OP_ERR;
 444         goto err;
 445
 446 acc_err:
 447         send_status = IB_WC_REM_ACCESS_ERR;
 448         wc.status = IB_WC_LOC_PROT_ERR;
 449 err:
 450         /* responder goes to error state */
 451         rvt_rc_error(qp, wc.status);
 452
 453 serr:
 454         spin_lock_irqsave(&sqp->s_lock, flags);
 455         qib_send_complete(sqp, wqe, send_status);
 456         if (sqp->ibqp.qp_type == IB_QPT_RC) {
 457                 int lastwqe = rvt_error_qp(sqp, IB_WC_WR_FLUSH_ERR);
 458
 459                 sqp->s_flags &= ~RVT_S_BUSY;
 460                 spin_unlock_irqrestore(&sqp->s_lock, flags);
 461                 if (lastwqe) {
 462                         struct ib_event ev;
 463
 464                         ev.device = sqp->ibqp.device;
 465                         ev.element.qp = &sqp->ibqp;
 466                         ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
 467                         sqp->ibqp.event_handler(&ev, sqp->ibqp.qp_context);
 468                 }
 469                 goto done;
 470         }
 471 clr_busy:
 472         sqp->s_flags &= ~RVT_S_BUSY;
 473 unlock:
 474         spin_unlock_irqrestore(&sqp->s_lock, flags);
 475 done:
 476         rcu_read_unlock();
 477 }
 478
 479 /**
 480  * qib_make_grh - construct a GRH header
 481  * @ibp: a pointer to the IB port
 482  * @hdr: a pointer to the GRH header being constructed
 483  * @grh: the global route address to send to
 484  * @hwords: the number of 32 bit words of header being sent
 485  * @nwords: the number of 32 bit words of data being sent
 486  *
 487  * Return the size of the header in 32 bit words.
 488  */
 489 u32 qib_make_grh(struct qib_ibport *ibp, struct ib_grh *hdr,
 490                  const struct ib_global_route *grh, u32 hwords, u32 nwords)
 491 {
 492         hdr->version_tclass_flow =
 493                 cpu_to_be32((IB_GRH_VERSION << IB_GRH_VERSION_SHIFT) |
 494                             (grh->traffic_class << IB_GRH_TCLASS_SHIFT) |
 495                             (grh->flow_label << IB_GRH_FLOW_SHIFT));
 496         hdr->paylen = cpu_to_be16((hwords - 2 + nwords + SIZE_OF_CRC) << 2);
 497         /* next_hdr is defined by C8-7 in ch. 8.4.1 */
 498         hdr->next_hdr = IB_GRH_NEXT_HDR;
 499         hdr->hop_limit = grh->hop_limit;
 500         /* The SGID is 32-bit aligned. */
 501         hdr->sgid.global.subnet_prefix = ibp->rvp.gid_prefix;
 502         if (!grh->sgid_index)
 503                 hdr->sgid.global.interface_id = ppd_from_ibp(ibp)->guid;
 504         else if (grh->sgid_index < QIB_GUIDS_PER_PORT)
 505                 hdr->sgid.global.interface_id = ibp->guids[grh->sgid_index - 1];
 506         hdr->dgid = grh->dgid;
 507
 508         /* GRH header size in 32-bit words. */
 509         return sizeof(struct ib_grh) / sizeof(u32);
 510 }
 511
 512 void qib_make_ruc_header(struct rvt_qp *qp, struct ib_other_headers *ohdr,
 513                          u32 bth0, u32 bth2)
 514 {
 515         struct qib_qp_priv *priv = qp->priv;
 516         struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
 517         u16 lrh0;
 518         u32 nwords;
 519         u32 extra_bytes;
 520
 521         /* Construct the header. */
 522         extra_bytes = -qp->s_cur_size & 3;
 523         nwords = (qp->s_cur_size + extra_bytes) >> 2;
 524         lrh0 = QIB_LRH_BTH;
 525         if (unlikely(rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH)) {
 526                 qp->s_hdrwords +=
 527                         qib_make_grh(ibp, &priv->s_hdr->u.l.grh,
 528                                      rdma_ah_read_grh(&qp->remote_ah_attr),
 529                                      qp->s_hdrwords, nwords);
 530                 lrh0 = QIB_LRH_GRH;
 531         }
 532         lrh0 |= ibp->sl_to_vl[rdma_ah_get_sl(&qp->remote_ah_attr)] << 12 |
 533                 rdma_ah_get_sl(&qp->remote_ah_attr) << 4;
 534         priv->s_hdr->lrh[0] = cpu_to_be16(lrh0);
 535         priv->s_hdr->lrh[1] =
 536                         cpu_to_be16(rdma_ah_get_dlid(&qp->remote_ah_attr));
 537         priv->s_hdr->lrh[2] =
 538                         cpu_to_be16(qp->s_hdrwords + nwords + SIZE_OF_CRC);
 539         priv->s_hdr->lrh[3] =
 540                 cpu_to_be16(ppd_from_ibp(ibp)->lid |
 541                             rdma_ah_get_path_bits(&qp->remote_ah_attr));
 542         bth0 |= qib_get_pkey(ibp, qp->s_pkey_index);
 543         bth0 |= extra_bytes << 20;
 544         if (qp->s_mig_state == IB_MIG_MIGRATED)
 545                 bth0 |= IB_BTH_MIG_REQ;
 546         ohdr->bth[0] = cpu_to_be32(bth0);
 547         ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
 548         ohdr->bth[2] = cpu_to_be32(bth2);
 549         this_cpu_inc(ibp->pmastats->n_unicast_xmit);
 550 }
 551
 552 void _qib_do_send(struct work_struct *work)
 553 {
 554         struct qib_qp_priv *priv = container_of(work, struct qib_qp_priv,
 555                                                 s_work);
 556         struct rvt_qp *qp = priv->owner;
 557
 558         qib_do_send(qp);
 559 }
 560
 561 /**
 562  * qib_do_send - perform a send on a QP
 563  * @qp: pointer to the QP
 564  *
 565  * Process entries in the send work queue until credit or queue is
 566  * exhausted.  Only allow one CPU to send a packet per QP (tasklet).
 567  * Otherwise, two threads could send packets out of order.
 568  */
 569 void qib_do_send(struct rvt_qp *qp)
 570 {
 571         struct qib_qp_priv *priv = qp->priv;
 572         struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
 573         struct qib_pportdata *ppd = ppd_from_ibp(ibp);
 574         int (*make_req)(struct rvt_qp *qp, unsigned long *flags);
 575         unsigned long flags;
 576
 577         if ((qp->ibqp.qp_type == IB_QPT_RC ||
 578              qp->ibqp.qp_type == IB_QPT_UC) &&
 579             (rdma_ah_get_dlid(&qp->remote_ah_attr) &
 580              ~((1 << ppd->lmc) - 1)) == ppd->lid) {
 581                 qib_ruc_loopback(qp);
 582                 return;
 583         }
 584
 585         if (qp->ibqp.qp_type == IB_QPT_RC)
 586                 make_req = qib_make_rc_req;
 587         else if (qp->ibqp.qp_type == IB_QPT_UC)
 588                 make_req = qib_make_uc_req;
 589         else
 590                 make_req = qib_make_ud_req;
 591
 592         spin_lock_irqsave(&qp->s_lock, flags);
 593
 594         /* Return if we are already busy processing a work request. */
 595         if (!qib_send_ok(qp)) {
 596                 spin_unlock_irqrestore(&qp->s_lock, flags);
 597                 return;
 598         }
 599
 600         qp->s_flags |= RVT_S_BUSY;
 601
 602         do {
 603                 /* Check for a constructed packet to be sent. */
 604                 if (qp->s_hdrwords != 0) {
 605                         spin_unlock_irqrestore(&qp->s_lock, flags);
 606                         /*
 607                          * If the packet cannot be sent now, return and
 608                          * the send tasklet will be woken up later.
 609                          */
 610                         if (qib_verbs_send(qp, priv->s_hdr, qp->s_hdrwords,
 611                                            qp->s_cur_sge, qp->s_cur_size))
 612                                 return;
 613                         /* Record that s_hdr is empty. */
 614                         qp->s_hdrwords = 0;
 615                         spin_lock_irqsave(&qp->s_lock, flags);
 616                 }
 617         } while (make_req(qp, &flags));
 618
 619         spin_unlock_irqrestore(&qp->s_lock, flags);
 620 }
 621
 622 /*
 623  * This should be called with s_lock held.
 624  */
 625 void qib_send_complete(struct rvt_qp *qp, struct rvt_swqe *wqe,
 626                        enum ib_wc_status status)
 627 {
 628         u32 old_last, last;
 629
 630         if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_OR_FLUSH_SEND))
 631                 return;
 632
 633         last = qp->s_last;
 634         old_last = last;
 635         if (++last >= qp->s_size)
 636                 last = 0;
 637         qp->s_last = last;
 638         /* See post_send() */
 639         barrier();
 640         rvt_put_swqe(wqe);
 641         if (qp->ibqp.qp_type == IB_QPT_UD ||
 642             qp->ibqp.qp_type == IB_QPT_SMI ||
 643             qp->ibqp.qp_type == IB_QPT_GSI)
 644                 atomic_dec(&ibah_to_rvtah(wqe->ud_wr.ah)->refcount);
 645
 646         rvt_qp_swqe_complete(qp,
 647                              wqe,
 648                              ib_qib_wc_opcode[wqe->wr.opcode],
 649                              status);
 650
 651         if (qp->s_acked == old_last)
 652                 qp->s_acked = last;
 653         if (qp->s_cur == old_last)
 654                 qp->s_cur = last;
 655         if (qp->s_tail == old_last)
 656                 qp->s_tail = last;
 657         if (qp->state == IB_QPS_SQD && last == qp->s_cur)
 658                 qp->s_draining = 0;
 659 }