git.oblomov.eu Git - linux-2.6/blob - drivers/infiniband/hw/ipath/ipath_ruc.c

   1 /*
   2  * Copyright (c) 2006 QLogic, Inc. 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 "ipath_verbs.h"
  35 #include "ipath_kernel.h"
  36
  37 /*
  38  * Convert the AETH RNR timeout code into the number of milliseconds.
  39  */
  40 const u32 ib_ipath_rnr_table[32] = {
  41         656,                    /* 0 */
  42         1,                      /* 1 */
  43         1,                      /* 2 */
  44         1,                      /* 3 */
  45         1,                      /* 4 */
  46         1,                      /* 5 */
  47         1,                      /* 6 */
  48         1,                      /* 7 */
  49         1,                      /* 8 */
  50         1,                      /* 9 */
  51         1,                      /* A */
  52         1,                      /* B */
  53         1,                      /* C */
  54         1,                      /* D */
  55         2,                      /* E */
  56         2,                      /* F */
  57         3,                      /* 10 */
  58         4,                      /* 11 */
  59         6,                      /* 12 */
  60         8,                      /* 13 */
  61         11,                     /* 14 */
  62         16,                     /* 15 */
  63         21,                     /* 16 */
  64         31,                     /* 17 */
  65         41,                     /* 18 */
  66         62,                     /* 19 */
  67         82,                     /* 1A */
  68         123,                    /* 1B */
  69         164,                    /* 1C */
  70         246,                    /* 1D */
  71         328,                    /* 1E */
  72         492                     /* 1F */
  73 };
  74
  75 /**
  76  * ipath_insert_rnr_queue - put QP on the RNR timeout list for the device
  77  * @qp: the QP
  78  *
  79  * XXX Use a simple list for now.  We might need a priority
  80  * queue if we have lots of QPs waiting for RNR timeouts
  81  * but that should be rare.
  82  */
  83 void ipath_insert_rnr_queue(struct ipath_qp *qp)
  84 {
  85         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
  86         unsigned long flags;
  87
  88         spin_lock_irqsave(&dev->pending_lock, flags);
  89         if (list_empty(&dev->rnrwait))
  90                 list_add(&qp->timerwait, &dev->rnrwait);
  91         else {
  92                 struct list_head *l = &dev->rnrwait;
  93                 struct ipath_qp *nqp = list_entry(l->next, struct ipath_qp,
  94                                                   timerwait);
  95
  96                 while (qp->s_rnr_timeout >= nqp->s_rnr_timeout) {
  97                         qp->s_rnr_timeout -= nqp->s_rnr_timeout;
  98                         l = l->next;
  99                         if (l->next == &dev->rnrwait)
 100                                 break;
 101                         nqp = list_entry(l->next, struct ipath_qp,
 102                                          timerwait);
 103                 }
 104                 list_add(&qp->timerwait, l);
 105         }
 106         spin_unlock_irqrestore(&dev->pending_lock, flags);
 107 }
 108
 109 static int init_sge(struct ipath_qp *qp, struct ipath_rwqe *wqe)
 110 {
 111         int user = to_ipd(qp->ibqp.pd)->user;
 112         int i, j, ret;
 113         struct ib_wc wc;
 114
 115         qp->r_len = 0;
 116         for (i = j = 0; i < wqe->num_sge; i++) {
 117                 if (wqe->sg_list[i].length == 0)
 118                         continue;
 119                 /* Check LKEY */
 120                 if ((user && wqe->sg_list[i].lkey == 0) ||
 121                     !ipath_lkey_ok(qp, &qp->r_sg_list[j], &wqe->sg_list[i],
 122                                    IB_ACCESS_LOCAL_WRITE))
 123                         goto bad_lkey;
 124                 qp->r_len += wqe->sg_list[i].length;
 125                 j++;
 126         }
 127         qp->r_sge.sge = qp->r_sg_list[0];
 128         qp->r_sge.sg_list = qp->r_sg_list + 1;
 129         qp->r_sge.num_sge = j;
 130         ret = 1;
 131         goto bail;
 132
 133 bad_lkey:
 134         wc.wr_id = wqe->wr_id;
 135         wc.status = IB_WC_LOC_PROT_ERR;
 136         wc.opcode = IB_WC_RECV;
 137         wc.vendor_err = 0;
 138         wc.byte_len = 0;
 139         wc.imm_data = 0;
 140         wc.qp_num = qp->ibqp.qp_num;
 141         wc.src_qp = 0;
 142         wc.wc_flags = 0;
 143         wc.pkey_index = 0;
 144         wc.slid = 0;
 145         wc.sl = 0;
 146         wc.dlid_path_bits = 0;
 147         wc.port_num = 0;
 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         if (qp->ibqp.srq) {
 176                 srq = to_isrq(qp->ibqp.srq);
 177                 handler = srq->ibsrq.event_handler;
 178                 rq = &srq->rq;
 179         } else {
 180                 srq = NULL;
 181                 handler = NULL;
 182                 rq = &qp->r_rq;
 183         }
 184
 185         spin_lock_irqsave(&rq->lock, flags);
 186         wq = rq->wq;
 187         tail = wq->tail;
 188         /* Validate tail before using it since it is user writable. */
 189         if (tail >= rq->size)
 190                 tail = 0;
 191         do {
 192                 if (unlikely(tail == wq->head)) {
 193                         spin_unlock_irqrestore(&rq->lock, flags);
 194                         ret = 0;
 195                         goto bail;
 196                 }
 197                 wqe = get_rwqe_ptr(rq, tail);
 198                 if (++tail >= rq->size)
 199                         tail = 0;
 200         } while (!wr_id_only && !init_sge(qp, wqe));
 201         qp->r_wr_id = wqe->wr_id;
 202         wq->tail = tail;
 203
 204         ret = 1;
 205         if (handler) {
 206                 u32 n;
 207
 208                 /*
 209                  * validate head pointer value and compute
 210                  * the number of remaining WQEs.
 211                  */
 212                 n = wq->head;
 213                 if (n >= rq->size)
 214                         n = 0;
 215                 if (n < tail)
 216                         n += rq->size - tail;
 217                 else
 218                         n -= tail;
 219                 if (n < srq->limit) {
 220                         struct ib_event ev;
 221
 222                         srq->limit = 0;
 223                         spin_unlock_irqrestore(&rq->lock, flags);
 224                         ev.device = qp->ibqp.device;
 225                         ev.element.srq = qp->ibqp.srq;
 226                         ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
 227                         handler(&ev, srq->ibsrq.srq_context);
 228                         goto bail;
 229                 }
 230         }
 231         spin_unlock_irqrestore(&rq->lock, flags);
 232         qp->r_wrid_valid = 1;
 233
 234 bail:
 235         return ret;
 236 }
 237
 238 /**
 239  * ipath_ruc_loopback - handle UC and RC lookback requests
 240  * @sqp: the loopback QP
 241  *
 242  * This is called from ipath_do_uc_send() or ipath_do_rc_send() to
 243  * forward a WQE addressed to the same HCA.
 244  * Note that although we are single threaded due to the tasklet, we still
 245  * have to protect against post_send().  We don't have to worry about
 246  * receive interrupts since this is a connected protocol and all packets
 247  * will pass through here.
 248  */
 249 static void ipath_ruc_loopback(struct ipath_qp *sqp)
 250 {
 251         struct ipath_ibdev *dev = to_idev(sqp->ibqp.device);
 252         struct ipath_qp *qp;
 253         struct ipath_swqe *wqe;
 254         struct ipath_sge *sge;
 255         unsigned long flags;
 256         struct ib_wc wc;
 257         u64 sdata;
 258
 259         qp = ipath_lookup_qpn(&dev->qp_table, sqp->remote_qpn);
 260         if (!qp) {
 261                 dev->n_pkt_drops++;
 262                 return;
 263         }
 264
 265 again:
 266         spin_lock_irqsave(&sqp->s_lock, flags);
 267
 268         if (!(ib_ipath_state_ops[sqp->state] & IPATH_PROCESS_SEND_OK)) {
 269                 spin_unlock_irqrestore(&sqp->s_lock, flags);
 270                 goto done;
 271         }
 272
 273         /* Get the next send request. */
 274         if (sqp->s_last == sqp->s_head) {
 275                 /* Send work queue is empty. */
 276                 spin_unlock_irqrestore(&sqp->s_lock, flags);
 277                 goto done;
 278         }
 279
 280         /*
 281          * We can rely on the entry not changing without the s_lock
 282          * being held until we update s_last.
 283          */
 284         wqe = get_swqe_ptr(sqp, sqp->s_last);
 285         spin_unlock_irqrestore(&sqp->s_lock, flags);
 286
 287         wc.wc_flags = 0;
 288         wc.imm_data = 0;
 289
 290         sqp->s_sge.sge = wqe->sg_list[0];
 291         sqp->s_sge.sg_list = wqe->sg_list + 1;
 292         sqp->s_sge.num_sge = wqe->wr.num_sge;
 293         sqp->s_len = wqe->length;
 294         switch (wqe->wr.opcode) {
 295         case IB_WR_SEND_WITH_IMM:
 296                 wc.wc_flags = IB_WC_WITH_IMM;
 297                 wc.imm_data = wqe->wr.imm_data;
 298                 /* FALLTHROUGH */
 299         case IB_WR_SEND:
 300                 if (!ipath_get_rwqe(qp, 0)) {
 301                 rnr_nak:
 302                         /* Handle RNR NAK */
 303                         if (qp->ibqp.qp_type == IB_QPT_UC)
 304                                 goto send_comp;
 305                         if (sqp->s_rnr_retry == 0) {
 306                                 wc.status = IB_WC_RNR_RETRY_EXC_ERR;
 307                                 goto err;
 308                         }
 309                         if (sqp->s_rnr_retry_cnt < 7)
 310                                 sqp->s_rnr_retry--;
 311                         dev->n_rnr_naks++;
 312                         sqp->s_rnr_timeout =
 313                                 ib_ipath_rnr_table[sqp->r_min_rnr_timer];
 314                         ipath_insert_rnr_queue(sqp);
 315                         goto done;
 316                 }
 317                 break;
 318
 319         case IB_WR_RDMA_WRITE_WITH_IMM:
 320                 wc.wc_flags = IB_WC_WITH_IMM;
 321                 wc.imm_data = wqe->wr.imm_data;
 322                 if (!ipath_get_rwqe(qp, 1))
 323                         goto rnr_nak;
 324                 /* FALLTHROUGH */
 325         case IB_WR_RDMA_WRITE:
 326                 if (wqe->length == 0)
 327                         break;
 328                 if (unlikely(!ipath_rkey_ok(qp, &qp->r_sge, wqe->length,
 329                                             wqe->wr.wr.rdma.remote_addr,
 330                                             wqe->wr.wr.rdma.rkey,
 331                                             IB_ACCESS_REMOTE_WRITE))) {
 332                 acc_err:
 333                         wc.status = IB_WC_REM_ACCESS_ERR;
 334                 err:
 335                         wc.wr_id = wqe->wr.wr_id;
 336                         wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
 337                         wc.vendor_err = 0;
 338                         wc.byte_len = 0;
 339                         wc.qp_num = sqp->ibqp.qp_num;
 340                         wc.src_qp = sqp->remote_qpn;
 341                         wc.pkey_index = 0;
 342                         wc.slid = sqp->remote_ah_attr.dlid;
 343                         wc.sl = sqp->remote_ah_attr.sl;
 344                         wc.dlid_path_bits = 0;
 345                         wc.port_num = 0;
 346                         ipath_sqerror_qp(sqp, &wc);
 347                         goto done;
 348                 }
 349                 break;
 350
 351         case IB_WR_RDMA_READ:
 352                 if (unlikely(!ipath_rkey_ok(qp, &sqp->s_sge, wqe->length,
 353                                             wqe->wr.wr.rdma.remote_addr,
 354                                             wqe->wr.wr.rdma.rkey,
 355                                             IB_ACCESS_REMOTE_READ)))
 356                         goto acc_err;
 357                 if (unlikely(!(qp->qp_access_flags &
 358                                IB_ACCESS_REMOTE_READ)))
 359                         goto acc_err;
 360                 qp->r_sge.sge = wqe->sg_list[0];
 361                 qp->r_sge.sg_list = wqe->sg_list + 1;
 362                 qp->r_sge.num_sge = wqe->wr.num_sge;
 363                 break;
 364
 365         case IB_WR_ATOMIC_CMP_AND_SWP:
 366         case IB_WR_ATOMIC_FETCH_AND_ADD:
 367                 if (unlikely(!ipath_rkey_ok(qp, &qp->r_sge, sizeof(u64),
 368                                             wqe->wr.wr.rdma.remote_addr,
 369                                             wqe->wr.wr.rdma.rkey,
 370                                             IB_ACCESS_REMOTE_ATOMIC)))
 371                         goto acc_err;
 372                 /* Perform atomic OP and save result. */
 373                 sdata = wqe->wr.wr.atomic.swap;
 374                 spin_lock_irqsave(&dev->pending_lock, flags);
 375                 qp->r_atomic_data = *(u64 *) qp->r_sge.sge.vaddr;
 376                 if (wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)
 377                         *(u64 *) qp->r_sge.sge.vaddr =
 378                                 qp->r_atomic_data + sdata;
 379                 else if (qp->r_atomic_data == wqe->wr.wr.atomic.compare_add)
 380                         *(u64 *) qp->r_sge.sge.vaddr = sdata;
 381                 spin_unlock_irqrestore(&dev->pending_lock, flags);
 382                 *(u64 *) sqp->s_sge.sge.vaddr = qp->r_atomic_data;
 383                 goto send_comp;
 384
 385         default:
 386                 goto done;
 387         }
 388
 389         sge = &sqp->s_sge.sge;
 390         while (sqp->s_len) {
 391                 u32 len = sqp->s_len;
 392
 393                 if (len > sge->length)
 394                         len = sge->length;
 395                 BUG_ON(len == 0);
 396                 ipath_copy_sge(&qp->r_sge, sge->vaddr, len);
 397                 sge->vaddr += len;
 398                 sge->length -= len;
 399                 sge->sge_length -= len;
 400                 if (sge->sge_length == 0) {
 401                         if (--sqp->s_sge.num_sge)
 402                                 *sge = *sqp->s_sge.sg_list++;
 403                 } else if (sge->length == 0 && sge->mr != NULL) {
 404                         if (++sge->n >= IPATH_SEGSZ) {
 405                                 if (++sge->m >= sge->mr->mapsz)
 406                                         break;
 407                                 sge->n = 0;
 408                         }
 409                         sge->vaddr =
 410                                 sge->mr->map[sge->m]->segs[sge->n].vaddr;
 411                         sge->length =
 412                                 sge->mr->map[sge->m]->segs[sge->n].length;
 413                 }
 414                 sqp->s_len -= len;
 415         }
 416
 417         if (wqe->wr.opcode == IB_WR_RDMA_WRITE ||
 418             wqe->wr.opcode == IB_WR_RDMA_READ)
 419                 goto send_comp;
 420
 421         if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM)
 422                 wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
 423         else
 424                 wc.opcode = IB_WC_RECV;
 425         wc.wr_id = qp->r_wr_id;
 426         wc.status = IB_WC_SUCCESS;
 427         wc.vendor_err = 0;
 428         wc.byte_len = wqe->length;
 429         wc.qp_num = qp->ibqp.qp_num;
 430         wc.src_qp = qp->remote_qpn;
 431         /* XXX do we know which pkey matched? Only needed for GSI. */
 432         wc.pkey_index = 0;
 433         wc.slid = qp->remote_ah_attr.dlid;
 434         wc.sl = qp->remote_ah_attr.sl;
 435         wc.dlid_path_bits = 0;
 436         /* Signal completion event if the solicited bit is set. */
 437         ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
 438                        wqe->wr.send_flags & IB_SEND_SOLICITED);
 439
 440 send_comp:
 441         sqp->s_rnr_retry = sqp->s_rnr_retry_cnt;
 442
 443         if (!test_bit(IPATH_S_SIGNAL_REQ_WR, &sqp->s_flags) ||
 444             (wqe->wr.send_flags & IB_SEND_SIGNALED)) {
 445                 wc.wr_id = wqe->wr.wr_id;
 446                 wc.status = IB_WC_SUCCESS;
 447                 wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
 448                 wc.vendor_err = 0;
 449                 wc.byte_len = wqe->length;
 450                 wc.qp_num = sqp->ibqp.qp_num;
 451                 wc.src_qp = 0;
 452                 wc.pkey_index = 0;
 453                 wc.slid = 0;
 454                 wc.sl = 0;
 455                 wc.dlid_path_bits = 0;
 456                 wc.port_num = 0;
 457                 ipath_cq_enter(to_icq(sqp->ibqp.send_cq), &wc, 0);
 458         }
 459
 460         /* Update s_last now that we are finished with the SWQE */
 461         spin_lock_irqsave(&sqp->s_lock, flags);
 462         if (++sqp->s_last >= sqp->s_size)
 463                 sqp->s_last = 0;
 464         spin_unlock_irqrestore(&sqp->s_lock, flags);
 465         goto again;
 466
 467 done:
 468         if (atomic_dec_and_test(&qp->refcount))
 469                 wake_up(&qp->wait);
 470 }
 471
 472 static int want_buffer(struct ipath_devdata *dd)
 473 {
 474         set_bit(IPATH_S_PIOINTBUFAVAIL, &dd->ipath_sendctrl);
 475         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
 476                          dd->ipath_sendctrl);
 477
 478         return 0;
 479 }
 480
 481 /**
 482  * ipath_no_bufs_available - tell the layer driver we need buffers
 483  * @qp: the QP that caused the problem
 484  * @dev: the device we ran out of buffers on
 485  *
 486  * Called when we run out of PIO buffers.
 487  */
 488 void ipath_no_bufs_available(struct ipath_qp *qp, struct ipath_ibdev *dev)
 489 {
 490         unsigned long flags;
 491
 492         spin_lock_irqsave(&dev->pending_lock, flags);
 493         if (list_empty(&qp->piowait))
 494                 list_add_tail(&qp->piowait, &dev->piowait);
 495         spin_unlock_irqrestore(&dev->pending_lock, flags);
 496         /*
 497          * Note that as soon as want_buffer() is called and
 498          * possibly before it returns, ipath_ib_piobufavail()
 499          * could be called.  If we are still in the tasklet function,
 500          * tasklet_hi_schedule() will not call us until the next time
 501          * tasklet_hi_schedule() is called.
 502          * We clear the tasklet flag now since we are committing to return
 503          * from the tasklet function.
 504          */
 505         clear_bit(IPATH_S_BUSY, &qp->s_flags);
 506         tasklet_unlock(&qp->s_task);
 507         want_buffer(dev->dd);
 508         dev->n_piowait++;
 509 }
 510
 511 /**
 512  * ipath_post_ruc_send - post RC and UC sends
 513  * @qp: the QP to post on
 514  * @wr: the work request to send
 515  */
 516 int ipath_post_ruc_send(struct ipath_qp *qp, struct ib_send_wr *wr)
 517 {
 518         struct ipath_swqe *wqe;
 519         unsigned long flags;
 520         u32 next;
 521         int i, j;
 522         int acc;
 523         int ret;
 524
 525         /*
 526          * Don't allow RDMA reads or atomic operations on UC or
 527          * undefined operations.
 528          * Make sure buffer is large enough to hold the result for atomics.
 529          */
 530         if (qp->ibqp.qp_type == IB_QPT_UC) {
 531                 if ((unsigned) wr->opcode >= IB_WR_RDMA_READ) {
 532                         ret = -EINVAL;
 533                         goto bail;
 534                 }
 535         } else if ((unsigned) wr->opcode > IB_WR_ATOMIC_FETCH_AND_ADD) {
 536                 ret = -EINVAL;
 537                 goto bail;
 538         } else if (wr->opcode >= IB_WR_ATOMIC_CMP_AND_SWP &&
 539                    (wr->num_sge == 0 ||
 540                     wr->sg_list[0].length < sizeof(u64) ||
 541                     wr->sg_list[0].addr & (sizeof(u64) - 1))) {
 542                 ret = -EINVAL;
 543                 goto bail;
 544         }
 545         /* IB spec says that num_sge == 0 is OK. */
 546         if (wr->num_sge > qp->s_max_sge) {
 547                 ret = -ENOMEM;
 548                 goto bail;
 549         }
 550         spin_lock_irqsave(&qp->s_lock, flags);
 551         next = qp->s_head + 1;
 552         if (next >= qp->s_size)
 553                 next = 0;
 554         if (next == qp->s_last) {
 555                 spin_unlock_irqrestore(&qp->s_lock, flags);
 556                 ret = -EINVAL;
 557                 goto bail;
 558         }
 559
 560         wqe = get_swqe_ptr(qp, qp->s_head);
 561         wqe->wr = *wr;
 562         wqe->ssn = qp->s_ssn++;
 563         wqe->sg_list[0].mr = NULL;
 564         wqe->sg_list[0].vaddr = NULL;
 565         wqe->sg_list[0].length = 0;
 566         wqe->sg_list[0].sge_length = 0;
 567         wqe->length = 0;
 568         acc = wr->opcode >= IB_WR_RDMA_READ ? IB_ACCESS_LOCAL_WRITE : 0;
 569         for (i = 0, j = 0; i < wr->num_sge; i++) {
 570                 if (to_ipd(qp->ibqp.pd)->user && wr->sg_list[i].lkey == 0) {
 571                         spin_unlock_irqrestore(&qp->s_lock, flags);
 572                         ret = -EINVAL;
 573                         goto bail;
 574                 }
 575                 if (wr->sg_list[i].length == 0)
 576                         continue;
 577                 if (!ipath_lkey_ok(qp, &wqe->sg_list[j], &wr->sg_list[i],
 578                                    acc)) {
 579                         spin_unlock_irqrestore(&qp->s_lock, flags);
 580                         ret = -EINVAL;
 581                         goto bail;
 582                 }
 583                 wqe->length += wr->sg_list[i].length;
 584                 j++;
 585         }
 586         wqe->wr.num_sge = j;
 587         qp->s_head = next;
 588         spin_unlock_irqrestore(&qp->s_lock, flags);
 589
 590         ipath_do_ruc_send((unsigned long) qp);
 591
 592         ret = 0;
 593
 594 bail:
 595         return ret;
 596 }
 597
 598 /**
 599  * ipath_make_grh - construct a GRH header
 600  * @dev: a pointer to the ipath device
 601  * @hdr: a pointer to the GRH header being constructed
 602  * @grh: the global route address to send to
 603  * @hwords: the number of 32 bit words of header being sent
 604  * @nwords: the number of 32 bit words of data being sent
 605  *
 606  * Return the size of the header in 32 bit words.
 607  */
 608 u32 ipath_make_grh(struct ipath_ibdev *dev, struct ib_grh *hdr,
 609                    struct ib_global_route *grh, u32 hwords, u32 nwords)
 610 {
 611         hdr->version_tclass_flow =
 612                 cpu_to_be32((6 << 28) |
 613                             (grh->traffic_class << 20) |
 614                             grh->flow_label);
 615         hdr->paylen = cpu_to_be16((hwords - 2 + nwords + SIZE_OF_CRC) << 2);
 616         /* next_hdr is defined by C8-7 in ch. 8.4.1 */
 617         hdr->next_hdr = 0x1B;
 618         hdr->hop_limit = grh->hop_limit;
 619         /* The SGID is 32-bit aligned. */
 620         hdr->sgid.global.subnet_prefix = dev->gid_prefix;
 621         hdr->sgid.global.interface_id = dev->dd->ipath_guid;
 622         hdr->dgid = grh->dgid;
 623
 624         /* GRH header size in 32-bit words. */
 625         return sizeof(struct ib_grh) / sizeof(u32);
 626 }
 627
 628 /**
 629  * ipath_do_ruc_send - perform a send on an RC or UC QP
 630  * @data: contains a pointer to the QP
 631  *
 632  * Process entries in the send work queue until credit or queue is
 633  * exhausted.  Only allow one CPU to send a packet per QP (tasklet).
 634  * Otherwise, after we drop the QP s_lock, two threads could send
 635  * packets out of order.
 636  */
 637 void ipath_do_ruc_send(unsigned long data)
 638 {
 639         struct ipath_qp *qp = (struct ipath_qp *)data;
 640         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
 641         unsigned long flags;
 642         u16 lrh0;
 643         u32 nwords;
 644         u32 extra_bytes;
 645         u32 bth0;
 646         u32 bth2;
 647         u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
 648         struct ipath_other_headers *ohdr;
 649
 650         if (test_and_set_bit(IPATH_S_BUSY, &qp->s_flags))
 651                 goto bail;
 652
 653         if (unlikely(qp->remote_ah_attr.dlid == dev->dd->ipath_lid)) {
 654                 ipath_ruc_loopback(qp);
 655                 goto clear;
 656         }
 657
 658         ohdr = &qp->s_hdr.u.oth;
 659         if (qp->remote_ah_attr.ah_flags & IB_AH_GRH)
 660                 ohdr = &qp->s_hdr.u.l.oth;
 661
 662 again:
 663         /* Check for a constructed packet to be sent. */
 664         if (qp->s_hdrwords != 0) {
 665                 /*
 666                  * If no PIO bufs are available, return.  An interrupt will
 667                  * call ipath_ib_piobufavail() when one is available.
 668                  */
 669                 if (ipath_verbs_send(dev->dd, qp->s_hdrwords,
 670                                      (u32 *) &qp->s_hdr, qp->s_cur_size,
 671                                      qp->s_cur_sge)) {
 672                         ipath_no_bufs_available(qp, dev);
 673                         goto bail;
 674                 }
 675                 dev->n_unicast_xmit++;
 676                 /* Record that we sent the packet and s_hdr is empty. */
 677                 qp->s_hdrwords = 0;
 678         }
 679
 680         /*
 681          * The lock is needed to synchronize between setting
 682          * qp->s_ack_state, resend timer, and post_send().
 683          */
 684         spin_lock_irqsave(&qp->s_lock, flags);
 685
 686         /* Sending responses has higher priority over sending requests. */
 687         if (qp->s_ack_state != IB_OPCODE_RC_ACKNOWLEDGE &&
 688             (bth0 = ipath_make_rc_ack(qp, ohdr, pmtu)) != 0)
 689                 bth2 = qp->s_ack_psn++ & IPATH_PSN_MASK;
 690         else if (!((qp->ibqp.qp_type == IB_QPT_RC) ?
 691                    ipath_make_rc_req(qp, ohdr, pmtu, &bth0, &bth2) :
 692                    ipath_make_uc_req(qp, ohdr, pmtu, &bth0, &bth2))) {
 693                 /*
 694                  * Clear the busy bit before unlocking to avoid races with
 695                  * adding new work queue items and then failing to process
 696                  * them.
 697                  */
 698                 clear_bit(IPATH_S_BUSY, &qp->s_flags);
 699                 spin_unlock_irqrestore(&qp->s_lock, flags);
 700                 goto bail;
 701         }
 702
 703         spin_unlock_irqrestore(&qp->s_lock, flags);
 704
 705         /* Construct the header. */
 706         extra_bytes = (4 - qp->s_cur_size) & 3;
 707         nwords = (qp->s_cur_size + extra_bytes) >> 2;
 708         lrh0 = IPATH_LRH_BTH;
 709         if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
 710                 qp->s_hdrwords += ipath_make_grh(dev, &qp->s_hdr.u.l.grh,
 711                                                  &qp->remote_ah_attr.grh,
 712                                                  qp->s_hdrwords, nwords);
 713                 lrh0 = IPATH_LRH_GRH;
 714         }
 715         lrh0 |= qp->remote_ah_attr.sl << 4;
 716         qp->s_hdr.lrh[0] = cpu_to_be16(lrh0);
 717         qp->s_hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
 718         qp->s_hdr.lrh[2] = cpu_to_be16(qp->s_hdrwords + nwords +
 719                                        SIZE_OF_CRC);
 720         qp->s_hdr.lrh[3] = cpu_to_be16(dev->dd->ipath_lid);
 721         bth0 |= ipath_get_pkey(dev->dd, qp->s_pkey_index);
 722         bth0 |= extra_bytes << 20;
 723         ohdr->bth[0] = cpu_to_be32(bth0);
 724         ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
 725         ohdr->bth[2] = cpu_to_be32(bth2);
 726
 727         /* Check for more work to do. */
 728         goto again;
 729
 730 clear:
 731         clear_bit(IPATH_S_BUSY, &qp->s_flags);
 732 bail:
 733         return;
 734 }