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