drivers/infiniband/hw/ipath/ipath_ruc.c

   1 /*
   2  * Copyright (c) 2006, 2007 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         wc.wr_id = wqe->wr_id;
 144         wc.status = IB_WC_LOC_PROT_ERR;
 145         wc.opcode = IB_WC_RECV;
 146         wc.vendor_err = 0;
 147         wc.byte_len = 0;
 148         wc.imm_data = 0;
 149         wc.qp = &qp->ibqp;
 150         wc.src_qp = 0;
 151         wc.wc_flags = 0;
 152         wc.pkey_index = 0;
 153         wc.slid = 0;
 154         wc.sl = 0;
 155         wc.dlid_path_bits = 0;
 156         wc.port_num = 0;
 157         /* Signal solicited completion event. */
 158         ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
 159         ret = 0;
 160 bail:
 161         return ret;
 162 }
 163
 164 /**
 165  * ipath_get_rwqe - copy the next RWQE into the QP's RWQE
 166  * @qp: the QP
 167  * @wr_id_only: update wr_id only, not SGEs
 168  *
 169  * Return 0 if no RWQE is available, otherwise return 1.
 170  *
 171  * Can be called from interrupt level.
 172  */
 173 int ipath_get_rwqe(struct ipath_qp *qp, int wr_id_only)
 174 {
 175         unsigned long flags;
 176         struct ipath_rq *rq;
 177         struct ipath_rwq *wq;
 178         struct ipath_srq *srq;
 179         struct ipath_rwqe *wqe;
 180         void (*handler)(struct ib_event *, void *);
 181         u32 tail;
 182         int ret;
 183
 184         qp->r_sge.sg_list = qp->r_sg_list;
 185
 186         if (qp->ibqp.srq) {
 187                 srq = to_isrq(qp->ibqp.srq);
 188                 handler = srq->ibsrq.event_handler;
 189                 rq = &srq->rq;
 190         } else {
 191                 srq = NULL;
 192                 handler = NULL;
 193                 rq = &qp->r_rq;
 194         }
 195
 196         spin_lock_irqsave(&rq->lock, flags);
 197         wq = rq->wq;
 198         tail = wq->tail;
 199         /* Validate tail before using it since it is user writable. */
 200         if (tail >= rq->size)
 201                 tail = 0;
 202         do {
 203                 if (unlikely(tail == wq->head)) {
 204                         spin_unlock_irqrestore(&rq->lock, flags);
 205                         ret = 0;
 206                         goto bail;
 207                 }
 208                 /* Make sure entry is read after head index is read. */
 209                 smp_rmb();
 210                 wqe = get_rwqe_ptr(rq, tail);
 211                 if (++tail >= rq->size)
 212                         tail = 0;
 213         } while (!wr_id_only && !ipath_init_sge(qp, wqe, &qp->r_len,
 214                                                 &qp->r_sge));
 215         qp->r_wr_id = wqe->wr_id;
 216         wq->tail = tail;
 217
 218         ret = 1;
 219         qp->r_wrid_valid = 1;
 220         if (handler) {
 221                 u32 n;
 222
 223                 /*
 224                  * validate head pointer value and compute
 225                  * the number of remaining WQEs.
 226                  */
 227                 n = wq->head;
 228                 if (n >= rq->size)
 229                         n = 0;
 230                 if (n < tail)
 231                         n += rq->size - tail;
 232                 else
 233                         n -= tail;
 234                 if (n < srq->limit) {
 235                         struct ib_event ev;
 236
 237                         srq->limit = 0;
 238                         spin_unlock_irqrestore(&rq->lock, flags);
 239                         ev.device = qp->ibqp.device;
 240                         ev.element.srq = qp->ibqp.srq;
 241                         ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
 242                         handler(&ev, srq->ibsrq.srq_context);
 243                         goto bail;
 244                 }
 245         }
 246         spin_unlock_irqrestore(&rq->lock, flags);
 247
 248 bail:
 249         return ret;
 250 }
 251
 252 /**
 253  * ipath_ruc_loopback - handle UC and RC lookback requests
 254  * @sqp: the sending QP
 255  *
 256  * This is called from ipath_do_send() to
 257  * forward a WQE addressed to the same HCA.
 258  * Note that although we are single threaded due to the tasklet, we still
 259  * have to protect against post_send().  We don't have to worry about
 260  * receive interrupts since this is a connected protocol and all packets
 261  * will pass through here.
 262  */
 263 static void ipath_ruc_loopback(struct ipath_qp *sqp)
 264 {
 265         struct ipath_ibdev *dev = to_idev(sqp->ibqp.device);
 266         struct ipath_qp *qp;
 267         struct ipath_swqe *wqe;
 268         struct ipath_sge *sge;
 269         unsigned long flags;
 270         struct ib_wc wc;
 271         u64 sdata;
 272         atomic64_t *maddr;
 273
 274         qp = ipath_lookup_qpn(&dev->qp_table, sqp->remote_qpn);
 275         if (!qp) {
 276                 dev->n_pkt_drops++;
 277                 return;
 278         }
 279
 280 again:
 281         spin_lock_irqsave(&sqp->s_lock, flags);
 282
 283         if (!(ib_ipath_state_ops[sqp->state] & IPATH_PROCESS_SEND_OK) ||
 284             sqp->s_rnr_timeout) {
 285                 spin_unlock_irqrestore(&sqp->s_lock, flags);
 286                 goto done;
 287         }
 288
 289         /* Get the next send request. */
 290         if (sqp->s_last == sqp->s_head) {
 291                 /* Send work queue is empty. */
 292                 spin_unlock_irqrestore(&sqp->s_lock, flags);
 293                 goto done;
 294         }
 295
 296         /*
 297          * We can rely on the entry not changing without the s_lock
 298          * being held until we update s_last.
 299          */
 300         wqe = get_swqe_ptr(sqp, sqp->s_last);
 301         spin_unlock_irqrestore(&sqp->s_lock, flags);
 302
 303         wc.wc_flags = 0;
 304         wc.imm_data = 0;
 305
 306         sqp->s_sge.sge = wqe->sg_list[0];
 307         sqp->s_sge.sg_list = wqe->sg_list + 1;
 308         sqp->s_sge.num_sge = wqe->wr.num_sge;
 309         sqp->s_len = wqe->length;
 310         switch (wqe->wr.opcode) {
 311         case IB_WR_SEND_WITH_IMM:
 312                 wc.wc_flags = IB_WC_WITH_IMM;
 313                 wc.imm_data = wqe->wr.imm_data;
 314                 /* FALLTHROUGH */
 315         case IB_WR_SEND:
 316                 if (!ipath_get_rwqe(qp, 0)) {
 317                 rnr_nak:
 318                         /* Handle RNR NAK */
 319                         if (qp->ibqp.qp_type == IB_QPT_UC)
 320                                 goto send_comp;
 321                         if (sqp->s_rnr_retry == 0) {
 322                                 wc.status = IB_WC_RNR_RETRY_EXC_ERR;
 323                                 goto err;
 324                         }
 325                         if (sqp->s_rnr_retry_cnt < 7)
 326                                 sqp->s_rnr_retry--;
 327                         dev->n_rnr_naks++;
 328                         sqp->s_rnr_timeout =
 329                                 ib_ipath_rnr_table[qp->r_min_rnr_timer];
 330                         ipath_insert_rnr_queue(sqp);
 331                         goto done;
 332                 }
 333                 break;
 334
 335         case IB_WR_RDMA_WRITE_WITH_IMM:
 336                 if (unlikely(!(qp->qp_access_flags &
 337                                IB_ACCESS_REMOTE_WRITE))) {
 338                         wc.status = IB_WC_REM_INV_REQ_ERR;
 339                         goto err;
 340                 }
 341                 wc.wc_flags = IB_WC_WITH_IMM;
 342                 wc.imm_data = wqe->wr.imm_data;
 343                 if (!ipath_get_rwqe(qp, 1))
 344                         goto rnr_nak;
 345                 /* FALLTHROUGH */
 346         case IB_WR_RDMA_WRITE:
 347                 if (unlikely(!(qp->qp_access_flags &
 348                                IB_ACCESS_REMOTE_WRITE))) {
 349                         wc.status = IB_WC_REM_INV_REQ_ERR;
 350                         goto err;
 351                 }
 352                 if (wqe->length == 0)
 353                         break;
 354                 if (unlikely(!ipath_rkey_ok(qp, &qp->r_sge, wqe->length,
 355                                             wqe->wr.wr.rdma.remote_addr,
 356                                             wqe->wr.wr.rdma.rkey,
 357                                             IB_ACCESS_REMOTE_WRITE))) {
 358                 acc_err:
 359                         wc.status = IB_WC_REM_ACCESS_ERR;
 360                 err:
 361                         wc.wr_id = wqe->wr.wr_id;
 362                         wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
 363                         wc.vendor_err = 0;
 364                         wc.byte_len = 0;
 365                         wc.qp = &sqp->ibqp;
 366                         wc.src_qp = sqp->remote_qpn;
 367                         wc.pkey_index = 0;
 368                         wc.slid = sqp->remote_ah_attr.dlid;
 369                         wc.sl = sqp->remote_ah_attr.sl;
 370                         wc.dlid_path_bits = 0;
 371                         wc.port_num = 0;
 372                         spin_lock_irqsave(&sqp->s_lock, flags);
 373                         ipath_sqerror_qp(sqp, &wc);
 374                         spin_unlock_irqrestore(&sqp->s_lock, flags);
 375                         goto done;
 376                 }
 377                 break;
 378
 379         case IB_WR_RDMA_READ:
 380                 if (unlikely(!(qp->qp_access_flags &
 381                                IB_ACCESS_REMOTE_READ))) {
 382                         wc.status = IB_WC_REM_INV_REQ_ERR;
 383                         goto err;
 384                 }
 385                 if (unlikely(!ipath_rkey_ok(qp, &sqp->s_sge, wqe->length,
 386                                             wqe->wr.wr.rdma.remote_addr,
 387                                             wqe->wr.wr.rdma.rkey,
 388                                             IB_ACCESS_REMOTE_READ)))
 389                         goto acc_err;
 390                 qp->r_sge.sge = wqe->sg_list[0];
 391                 qp->r_sge.sg_list = wqe->sg_list + 1;
 392                 qp->r_sge.num_sge = wqe->wr.num_sge;
 393                 break;
 394
 395         case IB_WR_ATOMIC_CMP_AND_SWP:
 396         case IB_WR_ATOMIC_FETCH_AND_ADD:
 397                 if (unlikely(!(qp->qp_access_flags &
 398                                IB_ACCESS_REMOTE_ATOMIC))) {
 399                         wc.status = IB_WC_REM_INV_REQ_ERR;
 400                         goto err;
 401                 }
 402                 if (unlikely(!ipath_rkey_ok(qp, &qp->r_sge, sizeof(u64),
 403                                             wqe->wr.wr.atomic.remote_addr,
 404                                             wqe->wr.wr.atomic.rkey,
 405                                             IB_ACCESS_REMOTE_ATOMIC)))
 406                         goto acc_err;
 407                 /* Perform atomic OP and save result. */
 408                 maddr = (atomic64_t *) qp->r_sge.sge.vaddr;
 409                 sdata = wqe->wr.wr.atomic.compare_add;
 410                 *(u64 *) sqp->s_sge.sge.vaddr =
 411                         (wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) ?
 412                         (u64) atomic64_add_return(sdata, maddr) - sdata :
 413                         (u64) cmpxchg((u64 *) qp->r_sge.sge.vaddr,
 414                                       sdata, wqe->wr.wr.atomic.swap);
 415                 goto send_comp;
 416
 417         default:
 418                 goto done;
 419         }
 420
 421         sge = &sqp->s_sge.sge;
 422         while (sqp->s_len) {
 423                 u32 len = sqp->s_len;
 424
 425                 if (len > sge->length)
 426                         len = sge->length;
 427                 if (len > sge->sge_length)
 428                         len = sge->sge_length;
 429                 BUG_ON(len == 0);
 430                 ipath_copy_sge(&qp->r_sge, sge->vaddr, len);
 431                 sge->vaddr += len;
 432                 sge->length -= len;
 433                 sge->sge_length -= len;
 434                 if (sge->sge_length == 0) {
 435                         if (--sqp->s_sge.num_sge)
 436                                 *sge = *sqp->s_sge.sg_list++;
 437                 } else if (sge->length == 0 && sge->mr != NULL) {
 438                         if (++sge->n >= IPATH_SEGSZ) {
 439                                 if (++sge->m >= sge->mr->mapsz)
 440                                         break;
 441                                 sge->n = 0;
 442                         }
 443                         sge->vaddr =
 444                                 sge->mr->map[sge->m]->segs[sge->n].vaddr;
 445                         sge->length =
 446                                 sge->mr->map[sge->m]->segs[sge->n].length;
 447                 }
 448                 sqp->s_len -= len;
 449         }
 450
 451         if (wqe->wr.opcode == IB_WR_RDMA_WRITE ||
 452             wqe->wr.opcode == IB_WR_RDMA_READ)
 453                 goto send_comp;
 454
 455         if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM)
 456                 wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
 457         else
 458                 wc.opcode = IB_WC_RECV;
 459         wc.wr_id = qp->r_wr_id;
 460         wc.status = IB_WC_SUCCESS;
 461         wc.vendor_err = 0;
 462         wc.byte_len = wqe->length;
 463         wc.qp = &qp->ibqp;
 464         wc.src_qp = qp->remote_qpn;
 465         wc.pkey_index = 0;
 466         wc.slid = qp->remote_ah_attr.dlid;
 467         wc.sl = qp->remote_ah_attr.sl;
 468         wc.dlid_path_bits = 0;
 469         wc.port_num = 1;
 470         /* Signal completion event if the solicited bit is set. */
 471         ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
 472                        wqe->wr.send_flags & IB_SEND_SOLICITED);
 473
 474 send_comp:
 475         sqp->s_rnr_retry = sqp->s_rnr_retry_cnt;
 476         ipath_send_complete(sqp, wqe, IB_WC_SUCCESS);
 477         goto again;
 478
 479 done:
 480         if (atomic_dec_and_test(&qp->refcount))
 481                 wake_up(&qp->wait);
 482 }
 483
 484 static void want_buffer(struct ipath_devdata *dd)
 485 {
 486         unsigned long flags;
 487
 488         spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
 489         dd->ipath_sendctrl |= INFINIPATH_S_PIOINTBUFAVAIL;
 490         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
 491                          dd->ipath_sendctrl);
 492         ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
 493         spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
 494 }
 495
 496 /**
 497  * ipath_no_bufs_available - tell the layer driver we need buffers
 498  * @qp: the QP that caused the problem
 499  * @dev: the device we ran out of buffers on
 500  *
 501  * Called when we run out of PIO buffers.
 502  */
 503 static void ipath_no_bufs_available(struct ipath_qp *qp,
 504                                     struct ipath_ibdev *dev)
 505 {
 506         unsigned long flags;
 507
 508         /*
 509          * Note that as soon as want_buffer() is called and
 510          * possibly before it returns, ipath_ib_piobufavail()
 511          * could be called.  If we are still in the tasklet function,
 512          * tasklet_hi_schedule() will not call us until the next time
 513          * tasklet_hi_schedule() is called.
 514          * We leave the busy flag set so that another post send doesn't
 515          * try to put the same QP on the piowait list again.
 516          */
 517         spin_lock_irqsave(&dev->pending_lock, flags);
 518         list_add_tail(&qp->piowait, &dev->piowait);
 519         spin_unlock_irqrestore(&dev->pending_lock, flags);
 520         want_buffer(dev->dd);
 521         dev->n_piowait++;
 522 }
 523
 524 /**
 525  * ipath_make_grh - construct a GRH header
 526  * @dev: a pointer to the ipath device
 527  * @hdr: a pointer to the GRH header being constructed
 528  * @grh: the global route address to send to
 529  * @hwords: the number of 32 bit words of header being sent
 530  * @nwords: the number of 32 bit words of data being sent
 531  *
 532  * Return the size of the header in 32 bit words.
 533  */
 534 u32 ipath_make_grh(struct ipath_ibdev *dev, struct ib_grh *hdr,
 535                    struct ib_global_route *grh, u32 hwords, u32 nwords)
 536 {
 537         hdr->version_tclass_flow =
 538                 cpu_to_be32((6 << 28) |
 539                             (grh->traffic_class << 20) |
 540                             grh->flow_label);
 541         hdr->paylen = cpu_to_be16((hwords - 2 + nwords + SIZE_OF_CRC) << 2);
 542         /* next_hdr is defined by C8-7 in ch. 8.4.1 */
 543         hdr->next_hdr = 0x1B;
 544         hdr->hop_limit = grh->hop_limit;
 545         /* The SGID is 32-bit aligned. */
 546         hdr->sgid.global.subnet_prefix = dev->gid_prefix;
 547         hdr->sgid.global.interface_id = dev->dd->ipath_guid;
 548         hdr->dgid = grh->dgid;
 549
 550         /* GRH header size in 32-bit words. */
 551         return sizeof(struct ib_grh) / sizeof(u32);
 552 }
 553
 554 void ipath_make_ruc_header(struct ipath_ibdev *dev, struct ipath_qp *qp,
 555                            struct ipath_other_headers *ohdr,
 556                            u32 bth0, u32 bth2)
 557 {
 558         u16 lrh0;
 559         u32 nwords;
 560         u32 extra_bytes;
 561
 562         /* Construct the header. */
 563         extra_bytes = -qp->s_cur_size & 3;
 564         nwords = (qp->s_cur_size + extra_bytes) >> 2;
 565         lrh0 = IPATH_LRH_BTH;
 566         if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
 567                 qp->s_hdrwords += ipath_make_grh(dev, &qp->s_hdr.u.l.grh,
 568                                                  &qp->remote_ah_attr.grh,
 569                                                  qp->s_hdrwords, nwords);
 570                 lrh0 = IPATH_LRH_GRH;
 571         }
 572         lrh0 |= qp->remote_ah_attr.sl << 4;
 573         qp->s_hdr.lrh[0] = cpu_to_be16(lrh0);
 574         qp->s_hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
 575         qp->s_hdr.lrh[2] = cpu_to_be16(qp->s_hdrwords + nwords + SIZE_OF_CRC);
 576         qp->s_hdr.lrh[3] = cpu_to_be16(dev->dd->ipath_lid);
 577         bth0 |= ipath_get_pkey(dev->dd, qp->s_pkey_index);
 578         bth0 |= extra_bytes << 20;
 579         ohdr->bth[0] = cpu_to_be32(bth0 | (1 << 22));
 580         ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
 581         ohdr->bth[2] = cpu_to_be32(bth2);
 582 }
 583
 584 /**
 585  * ipath_do_send - perform a send on a QP
 586  * @data: contains a pointer to the QP
 587  *
 588  * Process entries in the send work queue until credit or queue is
 589  * exhausted.  Only allow one CPU to send a packet per QP (tasklet).
 590  * Otherwise, two threads could send packets out of order.
 591  */
 592 void ipath_do_send(unsigned long data)
 593 {
 594         struct ipath_qp *qp = (struct ipath_qp *)data;
 595         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
 596         int (*make_req)(struct ipath_qp *qp);
 597
 598         if (test_and_set_bit(IPATH_S_BUSY, &qp->s_busy))
 599                 goto bail;
 600
 601         if ((qp->ibqp.qp_type == IB_QPT_RC ||
 602              qp->ibqp.qp_type == IB_QPT_UC) &&
 603             qp->remote_ah_attr.dlid == dev->dd->ipath_lid) {
 604                 ipath_ruc_loopback(qp);
 605                 goto clear;
 606         }
 607
 608         if (qp->ibqp.qp_type == IB_QPT_RC)
 609                make_req = ipath_make_rc_req;
 610         else if (qp->ibqp.qp_type == IB_QPT_UC)
 611                make_req = ipath_make_uc_req;
 612         else
 613                make_req = ipath_make_ud_req;
 614
 615 again:
 616         /* Check for a constructed packet to be sent. */
 617         if (qp->s_hdrwords != 0) {
 618                 /*
 619                  * If no PIO bufs are available, return.  An interrupt will
 620                  * call ipath_ib_piobufavail() when one is available.
 621                  */
 622                 if (ipath_verbs_send(qp, &qp->s_hdr, qp->s_hdrwords,
 623                                      qp->s_cur_sge, qp->s_cur_size)) {
 624                         ipath_no_bufs_available(qp, dev);
 625                         goto bail;
 626                 }
 627                 dev->n_unicast_xmit++;
 628                 /* Record that we sent the packet and s_hdr is empty. */
 629                 qp->s_hdrwords = 0;
 630         }
 631
 632         if (make_req(qp))
 633                 goto again;
 634 clear:
 635         clear_bit(IPATH_S_BUSY, &qp->s_busy);
 636 bail:;
 637 }
 638
 639 void ipath_send_complete(struct ipath_qp *qp, struct ipath_swqe *wqe,
 640                          enum ib_wc_status status)
 641 {
 642         unsigned long flags;
 643         u32 last;
 644
 645         /* See ch. 11.2.4.1 and 10.7.3.1 */
 646         if (!(qp->s_flags & IPATH_S_SIGNAL_REQ_WR) ||
 647             (wqe->wr.send_flags & IB_SEND_SIGNALED) ||
 648             status != IB_WC_SUCCESS) {
 649                 struct ib_wc wc;
 650
 651                 wc.wr_id = wqe->wr.wr_id;
 652                 wc.status = status;
 653                 wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
 654                 wc.vendor_err = 0;
 655                 wc.byte_len = wqe->length;
 656                 wc.imm_data = 0;
 657                 wc.qp = &qp->ibqp;
 658                 wc.src_qp = 0;
 659                 wc.wc_flags = 0;
 660                 wc.pkey_index = 0;
 661                 wc.slid = 0;
 662                 wc.sl = 0;
 663                 wc.dlid_path_bits = 0;
 664                 wc.port_num = 0;
 665                 ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 0);
 666         }
 667
 668         spin_lock_irqsave(&qp->s_lock, flags);
 669         last = qp->s_last;
 670         if (++last >= qp->s_size)
 671                 last = 0;
 672         qp->s_last = last;
 673         spin_unlock_irqrestore(&qp->s_lock, flags);
 674 }