err.no Git - linux-2.6/blob - drivers/infiniband/hw/ipath/ipath_ruc.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/spinlock.h>
  35
  36 #include "ipath_verbs.h"
  37 #include "ipath_kernel.h"
  38
  39 /*
  40  * Convert the AETH RNR timeout code into the number of milliseconds.
  41  */
  42 const u32 ib_ipath_rnr_table[32] = {
  43         656,                    /* 0 */
  44         1,                      /* 1 */
  45         1,                      /* 2 */
  46         1,                      /* 3 */
  47         1,                      /* 4 */
  48         1,                      /* 5 */
  49         1,                      /* 6 */
  50         1,                      /* 7 */
  51         1,                      /* 8 */
  52         1,                      /* 9 */
  53         1,                      /* A */
  54         1,                      /* B */
  55         1,                      /* C */
  56         1,                      /* D */
  57         2,                      /* E */
  58         2,                      /* F */
  59         3,                      /* 10 */
  60         4,                      /* 11 */
  61         6,                      /* 12 */
  62         8,                      /* 13 */
  63         11,                     /* 14 */
  64         16,                     /* 15 */
  65         21,                     /* 16 */
  66         31,                     /* 17 */
  67         41,                     /* 18 */
  68         62,                     /* 19 */
  69         82,                     /* 1A */
  70         123,                    /* 1B */
  71         164,                    /* 1C */
  72         246,                    /* 1D */
  73         328,                    /* 1E */
  74         492                     /* 1F */
  75 };
  76
  77 /**
  78  * ipath_insert_rnr_queue - put QP on the RNR timeout list for the device
  79  * @qp: the QP
  80  *
  81  * XXX Use a simple list for now.  We might need a priority
  82  * queue if we have lots of QPs waiting for RNR timeouts
  83  * but that should be rare.
  84  */
  85 void ipath_insert_rnr_queue(struct ipath_qp *qp)
  86 {
  87         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
  88         unsigned long flags;
  89
  90         spin_lock_irqsave(&dev->pending_lock, flags);
  91         if (list_empty(&dev->rnrwait))
  92                 list_add(&qp->timerwait, &dev->rnrwait);
  93         else {
  94                 struct list_head *l = &dev->rnrwait;
  95                 struct ipath_qp *nqp = list_entry(l->next, struct ipath_qp,
  96                                                   timerwait);
  97
  98                 while (qp->s_rnr_timeout >= nqp->s_rnr_timeout) {
  99                         qp->s_rnr_timeout -= nqp->s_rnr_timeout;
 100                         l = l->next;
 101                         if (l->next == &dev->rnrwait) {
 102                                 nqp = NULL;
 103                                 break;
 104                         }
 105                         nqp = list_entry(l->next, struct ipath_qp,
 106                                          timerwait);
 107                 }
 108                 if (nqp)
 109                         nqp->s_rnr_timeout -= qp->s_rnr_timeout;
 110                 list_add(&qp->timerwait, l);
 111         }
 112         spin_unlock_irqrestore(&dev->pending_lock, flags);
 113 }
 114
 115 /**
 116  * ipath_init_sge - Validate a RWQE and fill in the SGE state
 117  * @qp: the QP
 118  *
 119  * Return 1 if OK.
 120  */
 121 int ipath_init_sge(struct ipath_qp *qp, struct ipath_rwqe *wqe,
 122                    u32 *lengthp, struct ipath_sge_state *ss)
 123 {
 124         int i, j, ret;
 125         struct ib_wc wc;
 126
 127         *lengthp = 0;
 128         for (i = j = 0; i < wqe->num_sge; i++) {
 129                 if (wqe->sg_list[i].length == 0)
 130                         continue;
 131                 /* Check LKEY */
 132                 if (!ipath_lkey_ok(qp, j ? &ss->sg_list[j - 1] : &ss->sge,
 133                                    &wqe->sg_list[i], IB_ACCESS_LOCAL_WRITE))
 134                         goto bad_lkey;
 135                 *lengthp += wqe->sg_list[i].length;
 136                 j++;
 137         }
 138         ss->num_sge = j;
 139         ret = 1;
 140         goto bail;
 141
 142 bad_lkey:
 143         memset(&wc, 0, sizeof(wc));
 144         wc.wr_id = wqe->wr_id;
 145         wc.status = IB_WC_LOC_PROT_ERR;
 146         wc.opcode = IB_WC_RECV;
 147         wc.qp = &qp->ibqp;
 148         /* Signal solicited completion event. */
 149         ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
 150         ret = 0;
 151 bail:
 152         return ret;
 153 }
 154
 155 /**
 156  * ipath_get_rwqe - copy the next RWQE into the QP's RWQE
 157  * @qp: the QP
 158  * @wr_id_only: update wr_id only, not SGEs
 159  *
 160  * Return 0 if no RWQE is available, otherwise return 1.
 161  *
 162  * Can be called from interrupt level.
 163  */
 164 int ipath_get_rwqe(struct ipath_qp *qp, int wr_id_only)
 165 {
 166         unsigned long flags;
 167         struct ipath_rq *rq;
 168         struct ipath_rwq *wq;
 169         struct ipath_srq *srq;
 170         struct ipath_rwqe *wqe;
 171         void (*handler)(struct ib_event *, void *);
 172         u32 tail;
 173         int ret;
 174
 175         qp->r_sge.sg_list = qp->r_sg_list;
 176
 177         if (qp->ibqp.srq) {
 178                 srq = to_isrq(qp->ibqp.srq);
 179                 handler = srq->ibsrq.event_handler;
 180                 rq = &srq->rq;
 181         } else {
 182                 srq = NULL;
 183                 handler = NULL;
 184                 rq = &qp->r_rq;
 185         }
 186
 187         spin_lock_irqsave(&rq->lock, flags);
 188         wq = rq->wq;
 189         tail = wq->tail;
 190         /* Validate tail before using it since it is user writable. */
 191         if (tail >= rq->size)
 192                 tail = 0;
 193         do {
 194                 if (unlikely(tail == wq->head)) {
 195                         spin_unlock_irqrestore(&rq->lock, flags);
 196                         ret = 0;
 197                         goto bail;
 198                 }
 199                 /* Make sure entry is read after head index is read. */
 200                 smp_rmb();
 201                 wqe = get_rwqe_ptr(rq, tail);
 202                 if (++tail >= rq->size)
 203                         tail = 0;
 204         } while (!wr_id_only && !ipath_init_sge(qp, wqe, &qp->r_len,
 205                                                 &qp->r_sge));
 206         qp->r_wr_id = wqe->wr_id;
 207         wq->tail = tail;
 208
 209         ret = 1;
 210         qp->r_wrid_valid = 1;
 211         if (handler) {
 212                 u32 n;
 213
 214                 /*
 215                  * validate head pointer value and compute
 216                  * the number of remaining WQEs.
 217                  */
 218                 n = wq->head;
 219                 if (n >= rq->size)
 220                         n = 0;
 221                 if (n < tail)
 222                         n += rq->size - tail;
 223                 else
 224                         n -= tail;
 225                 if (n < srq->limit) {
 226                         struct ib_event ev;
 227
 228                         srq->limit = 0;
 229                         spin_unlock_irqrestore(&rq->lock, flags);
 230                         ev.device = qp->ibqp.device;
 231                         ev.element.srq = qp->ibqp.srq;
 232                         ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
 233                         handler(&ev, srq->ibsrq.srq_context);
 234                         goto bail;
 235                 }
 236         }
 237         spin_unlock_irqrestore(&rq->lock, flags);
 238
 239 bail:
 240         return ret;
 241 }
 242
 243 /**
 244  * ipath_ruc_loopback - handle UC and RC lookback requests
 245  * @sqp: the sending QP
 246  *
 247  * This is called from ipath_do_send() to
 248  * forward a WQE addressed to the same HCA.
 249  * Note that although we are single threaded due to the tasklet, we still
 250  * have to protect against post_send().  We don't have to worry about
 251  * receive interrupts since this is a connected protocol and all packets
 252  * will pass through here.
 253  */
 254 static void ipath_ruc_loopback(struct ipath_qp *sqp)
 255 {
 256         struct ipath_ibdev *dev = to_idev(sqp->ibqp.device);
 257         struct ipath_qp *qp;
 258         struct ipath_swqe *wqe;
 259         struct ipath_sge *sge;
 260         unsigned long flags;
 261         struct ib_wc wc;
 262         u64 sdata;
 263         atomic64_t *maddr;
 264         enum ib_wc_status send_status;
 265
 266         qp = ipath_lookup_qpn(&dev->qp_table, sqp->remote_qpn);
 267         if (!qp) {
 268                 dev->n_pkt_drops++;
 269                 return;
 270         }
 271
 272 again:
 273         spin_lock_irqsave(&sqp->s_lock, flags);
 274
 275         if (!(ib_ipath_state_ops[sqp->state] & IPATH_PROCESS_SEND_OK) ||
 276             sqp->s_rnr_timeout) {
 277                 spin_unlock_irqrestore(&sqp->s_lock, flags);
 278                 goto done;
 279         }
 280
 281         /* Get the next send request. */
 282         if (sqp->s_last == sqp->s_head) {
 283                 /* Send work queue is empty. */
 284                 spin_unlock_irqrestore(&sqp->s_lock, flags);
 285                 goto done;
 286         }
 287
 288         /*
 289          * We can rely on the entry not changing without the s_lock
 290          * being held until we update s_last.
 291          */
 292         wqe = get_swqe_ptr(sqp, sqp->s_last);
 293         spin_unlock_irqrestore(&sqp->s_lock, flags);
 294
 295         memset(&wc, 0, sizeof wc);
 296         send_status = IB_WC_SUCCESS;
 297
 298         sqp->s_sge.sge = wqe->sg_list[0];
 299         sqp->s_sge.sg_list = wqe->sg_list + 1;
 300         sqp->s_sge.num_sge = wqe->wr.num_sge;
 301         sqp->s_len = wqe->length;
 302         switch (wqe->wr.opcode) {
 303         case IB_WR_SEND_WITH_IMM:
 304                 wc.wc_flags = IB_WC_WITH_IMM;
 305                 wc.imm_data = wqe->wr.ex.imm_data;
 306                 /* FALLTHROUGH */
 307         case IB_WR_SEND:
 308                 if (!ipath_get_rwqe(qp, 0))
 309                         goto rnr_nak;
 310                 break;
 311
 312         case IB_WR_RDMA_WRITE_WITH_IMM:
 313                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
 314                         goto inv_err;
 315                 wc.wc_flags = IB_WC_WITH_IMM;
 316                 wc.imm_data = wqe->wr.ex.imm_data;
 317                 if (!ipath_get_rwqe(qp, 1))
 318                         goto rnr_nak;
 319                 /* FALLTHROUGH */
 320         case IB_WR_RDMA_WRITE:
 321                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
 322                         goto inv_err;
 323                 if (wqe->length == 0)
 324                         break;
 325                 if (unlikely(!ipath_rkey_ok(qp, &qp->r_sge, wqe->length,
 326                                             wqe->wr.wr.rdma.remote_addr,
 327                                             wqe->wr.wr.rdma.rkey,
 328                                             IB_ACCESS_REMOTE_WRITE)))
 329                         goto acc_err;
 330                 break;
 331
 332         case IB_WR_RDMA_READ:
 333                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ)))
 334                         goto inv_err;
 335                 if (unlikely(!ipath_rkey_ok(qp, &sqp->s_sge, wqe->length,
 336                                             wqe->wr.wr.rdma.remote_addr,
 337                                             wqe->wr.wr.rdma.rkey,
 338                                             IB_ACCESS_REMOTE_READ)))
 339                         goto acc_err;
 340                 qp->r_sge.sge = wqe->sg_list[0];
 341                 qp->r_sge.sg_list = wqe->sg_list + 1;
 342                 qp->r_sge.num_sge = wqe->wr.num_sge;
 343                 break;
 344
 345         case IB_WR_ATOMIC_CMP_AND_SWP:
 346         case IB_WR_ATOMIC_FETCH_AND_ADD:
 347                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC)))
 348                         goto inv_err;
 349                 if (unlikely(!ipath_rkey_ok(qp, &qp->r_sge, sizeof(u64),
 350                                             wqe->wr.wr.atomic.remote_addr,
 351                                             wqe->wr.wr.atomic.rkey,
 352                                             IB_ACCESS_REMOTE_ATOMIC)))
 353                         goto acc_err;
 354                 /* Perform atomic OP and save result. */
 355                 maddr = (atomic64_t *) qp->r_sge.sge.vaddr;
 356                 sdata = wqe->wr.wr.atomic.compare_add;
 357                 *(u64 *) sqp->s_sge.sge.vaddr =
 358                         (wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) ?
 359                         (u64) atomic64_add_return(sdata, maddr) - sdata :
 360                         (u64) cmpxchg((u64 *) qp->r_sge.sge.vaddr,
 361                                       sdata, wqe->wr.wr.atomic.swap);
 362                 goto send_comp;
 363
 364         default:
 365                 send_status = IB_WC_LOC_QP_OP_ERR;
 366                 goto serr;
 367         }
 368
 369         sge = &sqp->s_sge.sge;
 370         while (sqp->s_len) {
 371                 u32 len = sqp->s_len;
 372
 373                 if (len > sge->length)
 374                         len = sge->length;
 375                 if (len > sge->sge_length)
 376                         len = sge->sge_length;
 377                 BUG_ON(len == 0);
 378                 ipath_copy_sge(&qp->r_sge, sge->vaddr, len);
 379                 sge->vaddr += len;
 380                 sge->length -= len;
 381                 sge->sge_length -= len;
 382                 if (sge->sge_length == 0) {
 383                         if (--sqp->s_sge.num_sge)
 384                                 *sge = *sqp->s_sge.sg_list++;
 385                 } else if (sge->length == 0 && sge->mr != NULL) {
 386                         if (++sge->n >= IPATH_SEGSZ) {
 387                                 if (++sge->m >= sge->mr->mapsz)
 388                                         break;
 389                                 sge->n = 0;
 390                         }
 391                         sge->vaddr =
 392                                 sge->mr->map[sge->m]->segs[sge->n].vaddr;
 393                         sge->length =
 394                                 sge->mr->map[sge->m]->segs[sge->n].length;
 395                 }
 396                 sqp->s_len -= len;
 397         }
 398
 399         if (wqe->wr.opcode == IB_WR_RDMA_WRITE ||
 400             wqe->wr.opcode == IB_WR_RDMA_READ)
 401                 goto send_comp;
 402
 403         if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM)
 404                 wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
 405         else
 406                 wc.opcode = IB_WC_RECV;
 407         wc.wr_id = qp->r_wr_id;
 408         wc.status = IB_WC_SUCCESS;
 409         wc.byte_len = wqe->length;
 410         wc.qp = &qp->ibqp;
 411         wc.src_qp = qp->remote_qpn;
 412         wc.slid = qp->remote_ah_attr.dlid;
 413         wc.sl = qp->remote_ah_attr.sl;
 414         wc.port_num = 1;
 415         /* Signal completion event if the solicited bit is set. */
 416         ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
 417                        wqe->wr.send_flags & IB_SEND_SOLICITED);
 418
 419 send_comp:
 420         sqp->s_rnr_retry = sqp->s_rnr_retry_cnt;
 421         ipath_send_complete(sqp, wqe, send_status);
 422         goto again;
 423
 424 rnr_nak:
 425         /* Handle RNR NAK */
 426         if (qp->ibqp.qp_type == IB_QPT_UC)
 427                 goto send_comp;
 428         /*
 429          * Note: we don't need the s_lock held since the BUSY flag
 430          * makes this single threaded.
 431          */
 432         if (sqp->s_rnr_retry == 0) {
 433                 send_status = IB_WC_RNR_RETRY_EXC_ERR;
 434                 goto serr;
 435         }
 436         if (sqp->s_rnr_retry_cnt < 7)
 437                 sqp->s_rnr_retry--;
 438         spin_lock_irqsave(&sqp->s_lock, flags);
 439         if (!(ib_ipath_state_ops[sqp->state] & IPATH_PROCESS_RECV_OK))
 440                 goto unlock;
 441         dev->n_rnr_naks++;
 442         sqp->s_rnr_timeout = ib_ipath_rnr_table[qp->r_min_rnr_timer];
 443         ipath_insert_rnr_queue(sqp);
 444         goto unlock;
 445
 446 inv_err:
 447         send_status = IB_WC_REM_INV_REQ_ERR;
 448         wc.status = IB_WC_LOC_QP_OP_ERR;
 449         goto err;
 450
 451 acc_err:
 452         send_status = IB_WC_REM_ACCESS_ERR;
 453         wc.status = IB_WC_LOC_PROT_ERR;
 454 err:
 455         /* responder goes to error state */
 456         ipath_rc_error(qp, wc.status);
 457
 458 serr:
 459         spin_lock_irqsave(&sqp->s_lock, flags);
 460         ipath_send_complete(sqp, wqe, send_status);
 461         if (sqp->ibqp.qp_type == IB_QPT_RC) {
 462                 int lastwqe = ipath_error_qp(sqp, IB_WC_WR_FLUSH_ERR);
 463
 464                 sqp->s_flags &= ~IPATH_S_BUSY;
 465                 spin_unlock_irqrestore(&sqp->s_lock, flags);
 466                 if (lastwqe) {
 467                         struct ib_event ev;
 468
 469                         ev.device = sqp->ibqp.device;
 470                         ev.element.qp = &sqp->ibqp;
 471                         ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
 472                         sqp->ibqp.event_handler(&ev, sqp->ibqp.qp_context);
 473                 }
 474                 goto done;
 475         }
 476 unlock:
 477         spin_unlock_irqrestore(&sqp->s_lock, flags);
 478 done:
 479         if (atomic_dec_and_test(&qp->refcount))
 480                 wake_up(&qp->wait);
 481 }
 482
 483 static void want_buffer(struct ipath_devdata *dd, struct ipath_qp *qp)
 484 {
 485         if (!(dd->ipath_flags & IPATH_HAS_SEND_DMA) ||
 486                 qp->ibqp.qp_type == IB_QPT_SMI) {
 487                 unsigned long flags;
 488
 489                 spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
 490                 dd->ipath_sendctrl |= INFINIPATH_S_PIOINTBUFAVAIL;
 491                 ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
 492                                  dd->ipath_sendctrl);
 493                 ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
 494                 spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
 495         }
 496 }
 497
 498 /**
 499  * ipath_no_bufs_available - tell the layer driver we need buffers
 500  * @qp: the QP that caused the problem
 501  * @dev: the device we ran out of buffers on
 502  *
 503  * Called when we run out of PIO buffers.
 504  */
 505 static void ipath_no_bufs_available(struct ipath_qp *qp,
 506                                     struct ipath_ibdev *dev)
 507 {
 508         unsigned long flags;
 509
 510         /*
 511          * Note that as soon as want_buffer() is called and
 512          * possibly before it returns, ipath_ib_piobufavail()
 513          * could be called.  If we are still in the tasklet function,
 514          * tasklet_hi_schedule() will not call us until the next time
 515          * tasklet_hi_schedule() is called.
 516          * We leave the busy flag set so that another post send doesn't
 517          * try to put the same QP on the piowait list again.
 518          */
 519         spin_lock_irqsave(&dev->pending_lock, flags);
 520         list_add_tail(&qp->piowait, &dev->piowait);
 521         spin_unlock_irqrestore(&dev->pending_lock, flags);
 522         want_buffer(dev->dd, qp);
 523         dev->n_piowait++;
 524 }
 525
 526 /**
 527  * ipath_make_grh - construct a GRH header
 528  * @dev: a pointer to the ipath device
 529  * @hdr: a pointer to the GRH header being constructed
 530  * @grh: the global route address to send to
 531  * @hwords: the number of 32 bit words of header being sent
 532  * @nwords: the number of 32 bit words of data being sent
 533  *
 534  * Return the size of the header in 32 bit words.
 535  */
 536 u32 ipath_make_grh(struct ipath_ibdev *dev, struct ib_grh *hdr,
 537                    struct ib_global_route *grh, u32 hwords, u32 nwords)
 538 {
 539         hdr->version_tclass_flow =
 540                 cpu_to_be32((6 << 28) |
 541                             (grh->traffic_class << 20) |
 542                             grh->flow_label);
 543         hdr->paylen = cpu_to_be16((hwords - 2 + nwords + SIZE_OF_CRC) << 2);
 544         /* next_hdr is defined by C8-7 in ch. 8.4.1 */
 545         hdr->next_hdr = 0x1B;
 546         hdr->hop_limit = grh->hop_limit;
 547         /* The SGID is 32-bit aligned. */
 548         hdr->sgid.global.subnet_prefix = dev->gid_prefix;
 549         hdr->sgid.global.interface_id = dev->dd->ipath_guid;
 550         hdr->dgid = grh->dgid;
 551
 552         /* GRH header size in 32-bit words. */
 553         return sizeof(struct ib_grh) / sizeof(u32);
 554 }
 555
 556 void ipath_make_ruc_header(struct ipath_ibdev *dev, struct ipath_qp *qp,
 557                            struct ipath_other_headers *ohdr,
 558                            u32 bth0, u32 bth2)
 559 {
 560         u16 lrh0;
 561         u32 nwords;
 562         u32 extra_bytes;
 563
 564         /* Construct the header. */
 565         extra_bytes = -qp->s_cur_size & 3;
 566         nwords = (qp->s_cur_size + extra_bytes) >> 2;
 567         lrh0 = IPATH_LRH_BTH;
 568         if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
 569                 qp->s_hdrwords += ipath_make_grh(dev, &qp->s_hdr.u.l.grh,
 570                                                  &qp->remote_ah_attr.grh,
 571                                                  qp->s_hdrwords, nwords);
 572                 lrh0 = IPATH_LRH_GRH;
 573         }
 574         lrh0 |= qp->remote_ah_attr.sl << 4;
 575         qp->s_hdr.lrh[0] = cpu_to_be16(lrh0);
 576         qp->s_hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
 577         qp->s_hdr.lrh[2] = cpu_to_be16(qp->s_hdrwords + nwords + SIZE_OF_CRC);
 578         qp->s_hdr.lrh[3] = cpu_to_be16(dev->dd->ipath_lid);
 579         bth0 |= ipath_get_pkey(dev->dd, qp->s_pkey_index);
 580         bth0 |= extra_bytes << 20;
 581         ohdr->bth[0] = cpu_to_be32(bth0 | (1 << 22));
 582         ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
 583         ohdr->bth[2] = cpu_to_be32(bth2);
 584 }
 585
 586 /**
 587  * ipath_do_send - perform a send on a QP
 588  * @data: contains a pointer to the QP
 589  *
 590  * Process entries in the send work queue until credit or queue is
 591  * exhausted.  Only allow one CPU to send a packet per QP (tasklet).
 592  * Otherwise, two threads could send packets out of order.
 593  */
 594 void ipath_do_send(unsigned long data)
 595 {
 596         struct ipath_qp *qp = (struct ipath_qp *)data;
 597         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
 598         int (*make_req)(struct ipath_qp *qp);
 599
 600         if (test_and_set_bit(IPATH_S_BUSY, &qp->s_busy))
 601                 goto bail;
 602
 603         if ((qp->ibqp.qp_type == IB_QPT_RC ||
 604              qp->ibqp.qp_type == IB_QPT_UC) &&
 605             qp->remote_ah_attr.dlid == dev->dd->ipath_lid) {
 606                 ipath_ruc_loopback(qp);
 607                 goto clear;
 608         }
 609
 610         if (qp->ibqp.qp_type == IB_QPT_RC)
 611                make_req = ipath_make_rc_req;
 612         else if (qp->ibqp.qp_type == IB_QPT_UC)
 613                make_req = ipath_make_uc_req;
 614         else
 615                make_req = ipath_make_ud_req;
 616
 617 again:
 618         /* Check for a constructed packet to be sent. */
 619         if (qp->s_hdrwords != 0) {
 620                 /*
 621                  * If no PIO bufs are available, return.  An interrupt will
 622                  * call ipath_ib_piobufavail() when one is available.
 623                  */
 624                 if (ipath_verbs_send(qp, &qp->s_hdr, qp->s_hdrwords,
 625                                      qp->s_cur_sge, qp->s_cur_size)) {
 626                         ipath_no_bufs_available(qp, dev);
 627                         goto bail;
 628                 }
 629                 dev->n_unicast_xmit++;
 630                 /* Record that we sent the packet and s_hdr is empty. */
 631                 qp->s_hdrwords = 0;
 632         }
 633
 634         if (make_req(qp))
 635                 goto again;
 636 clear:
 637         clear_bit(IPATH_S_BUSY, &qp->s_busy);
 638 bail:;
 639 }
 640
 641 void ipath_send_complete(struct ipath_qp *qp, struct ipath_swqe *wqe,
 642                          enum ib_wc_status status)
 643 {
 644         unsigned long flags;
 645         u32 last;
 646
 647         /* See ch. 11.2.4.1 and 10.7.3.1 */
 648         if (!(qp->s_flags & IPATH_S_SIGNAL_REQ_WR) ||
 649             (wqe->wr.send_flags & IB_SEND_SIGNALED) ||
 650             status != IB_WC_SUCCESS) {
 651                 struct ib_wc wc;
 652
 653                 memset(&wc, 0, sizeof wc);
 654                 wc.wr_id = wqe->wr.wr_id;
 655                 wc.status = status;
 656                 wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
 657                 wc.qp = &qp->ibqp;
 658                 if (status == IB_WC_SUCCESS)
 659                         wc.byte_len = wqe->length;
 660                 ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc,
 661                                status != IB_WC_SUCCESS);
 662         }
 663
 664         spin_lock_irqsave(&qp->s_lock, flags);
 665         last = qp->s_last;
 666         if (++last >= qp->s_size)
 667                 last = 0;
 668         qp->s_last = last;
 669         spin_unlock_irqrestore(&qp->s_lock, flags);
 670 }