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 = &qp->ibqp;
 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         qp->r_wrid_valid = 1;
 206         if (handler) {
 207                 u32 n;
 208
 209                 /*
 210                  * validate head pointer value and compute
 211                  * the number of remaining WQEs.
 212                  */
 213                 n = wq->head;
 214                 if (n >= rq->size)
 215                         n = 0;
 216                 if (n < tail)
 217                         n += rq->size - tail;
 218                 else
 219                         n -= tail;
 220                 if (n < srq->limit) {
 221                         struct ib_event ev;
 222
 223                         srq->limit = 0;
 224                         spin_unlock_irqrestore(&rq->lock, flags);
 225                         ev.device = qp->ibqp.device;
 226                         ev.element.srq = qp->ibqp.srq;
 227                         ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
 228                         handler(&ev, srq->ibsrq.srq_context);
 229                         goto bail;
 230                 }
 231         }
 232         spin_unlock_irqrestore(&rq->lock, flags);
 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         atomic64_t *maddr;
 259
 260         qp = ipath_lookup_qpn(&dev->qp_table, sqp->remote_qpn);
 261         if (!qp) {
 262                 dev->n_pkt_drops++;
 263                 return;
 264         }
 265
 266 again:
 267         spin_lock_irqsave(&sqp->s_lock, flags);
 268
 269         if (!(ib_ipath_state_ops[sqp->state] & IPATH_PROCESS_SEND_OK) ||
 270             qp->s_rnr_timeout) {
 271                 spin_unlock_irqrestore(&sqp->s_lock, flags);
 272                 goto done;
 273         }
 274
 275         /* Get the next send request. */
 276         if (sqp->s_last == sqp->s_head) {
 277                 /* Send work queue is empty. */
 278                 spin_unlock_irqrestore(&sqp->s_lock, flags);
 279                 goto done;
 280         }
 281
 282         /*
 283          * We can rely on the entry not changing without the s_lock
 284          * being held until we update s_last.
 285          */
 286         wqe = get_swqe_ptr(sqp, sqp->s_last);
 287         spin_unlock_irqrestore(&sqp->s_lock, flags);
 288
 289         wc.wc_flags = 0;
 290         wc.imm_data = 0;
 291
 292         sqp->s_sge.sge = wqe->sg_list[0];
 293         sqp->s_sge.sg_list = wqe->sg_list + 1;
 294         sqp->s_sge.num_sge = wqe->wr.num_sge;
 295         sqp->s_len = wqe->length;
 296         switch (wqe->wr.opcode) {
 297         case IB_WR_SEND_WITH_IMM:
 298                 wc.wc_flags = IB_WC_WITH_IMM;
 299                 wc.imm_data = wqe->wr.imm_data;
 300                 /* FALLTHROUGH */
 301         case IB_WR_SEND:
 302                 if (!ipath_get_rwqe(qp, 0)) {
 303                 rnr_nak:
 304                         /* Handle RNR NAK */
 305                         if (qp->ibqp.qp_type == IB_QPT_UC)
 306                                 goto send_comp;
 307                         if (sqp->s_rnr_retry == 0) {
 308                                 wc.status = IB_WC_RNR_RETRY_EXC_ERR;
 309                                 goto err;
 310                         }
 311                         if (sqp->s_rnr_retry_cnt < 7)
 312                                 sqp->s_rnr_retry--;
 313                         dev->n_rnr_naks++;
 314                         sqp->s_rnr_timeout =
 315                                 ib_ipath_rnr_table[qp->r_min_rnr_timer];
 316                         ipath_insert_rnr_queue(sqp);
 317                         goto done;
 318                 }
 319                 break;
 320
 321         case IB_WR_RDMA_WRITE_WITH_IMM:
 322                 wc.wc_flags = IB_WC_WITH_IMM;
 323                 wc.imm_data = wqe->wr.imm_data;
 324                 if (!ipath_get_rwqe(qp, 1))
 325                         goto rnr_nak;
 326                 /* FALLTHROUGH */
 327         case IB_WR_RDMA_WRITE:
 328                 if (wqe->length == 0)
 329                         break;
 330                 if (unlikely(!ipath_rkey_ok(qp, &qp->r_sge, wqe->length,
 331                                             wqe->wr.wr.rdma.remote_addr,
 332                                             wqe->wr.wr.rdma.rkey,
 333                                             IB_ACCESS_REMOTE_WRITE))) {
 334                 acc_err:
 335                         wc.status = IB_WC_REM_ACCESS_ERR;
 336                 err:
 337                         wc.wr_id = wqe->wr.wr_id;
 338                         wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
 339                         wc.vendor_err = 0;
 340                         wc.byte_len = 0;
 341                         wc.qp = &sqp->ibqp;
 342                         wc.src_qp = sqp->remote_qpn;
 343                         wc.pkey_index = 0;
 344                         wc.slid = sqp->remote_ah_attr.dlid;
 345                         wc.sl = sqp->remote_ah_attr.sl;
 346                         wc.dlid_path_bits = 0;
 347                         wc.port_num = 0;
 348                         spin_lock_irqsave(&sqp->s_lock, flags);
 349                         ipath_sqerror_qp(sqp, &wc);
 350                         spin_unlock_irqrestore(&sqp->s_lock, flags);
 351                         goto done;
 352                 }
 353                 break;
 354
 355         case IB_WR_RDMA_READ:
 356                 if (unlikely(!(qp->qp_access_flags &
 357                                IB_ACCESS_REMOTE_READ)))
 358                         goto acc_err;
 359                 if (unlikely(!ipath_rkey_ok(qp, &sqp->s_sge, wqe->length,
 360                                             wqe->wr.wr.rdma.remote_addr,
 361                                             wqe->wr.wr.rdma.rkey,
 362                                             IB_ACCESS_REMOTE_READ)))
 363                         goto acc_err;
 364                 qp->r_sge.sge = wqe->sg_list[0];
 365                 qp->r_sge.sg_list = wqe->sg_list + 1;
 366                 qp->r_sge.num_sge = wqe->wr.num_sge;
 367                 break;
 368
 369         case IB_WR_ATOMIC_CMP_AND_SWP:
 370         case IB_WR_ATOMIC_FETCH_AND_ADD:
 371                 if (unlikely(!(qp->qp_access_flags &
 372                                IB_ACCESS_REMOTE_ATOMIC)))
 373                         goto acc_err;
 374                 if (unlikely(!ipath_rkey_ok(qp, &qp->r_sge, sizeof(u64),
 375                                             wqe->wr.wr.atomic.remote_addr,
 376                                             wqe->wr.wr.atomic.rkey,
 377                                             IB_ACCESS_REMOTE_ATOMIC)))
 378                         goto acc_err;
 379                 /* Perform atomic OP and save result. */
 380                 maddr = (atomic64_t *) qp->r_sge.sge.vaddr;
 381                 sdata = wqe->wr.wr.atomic.compare_add;
 382                 *(u64 *) sqp->s_sge.sge.vaddr =
 383                         (wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) ?
 384                         (u64) atomic64_add_return(sdata, maddr) - sdata :
 385                         (u64) cmpxchg((u64 *) qp->r_sge.sge.vaddr,
 386                                       sdata, wqe->wr.wr.atomic.swap);
 387                 goto send_comp;
 388
 389         default:
 390                 goto done;
 391         }
 392
 393         sge = &sqp->s_sge.sge;
 394         while (sqp->s_len) {
 395                 u32 len = sqp->s_len;
 396
 397                 if (len > sge->length)
 398                         len = sge->length;
 399                 BUG_ON(len == 0);
 400                 ipath_copy_sge(&qp->r_sge, sge->vaddr, len);
 401                 sge->vaddr += len;
 402                 sge->length -= len;
 403                 sge->sge_length -= len;
 404                 if (sge->sge_length == 0) {
 405                         if (--sqp->s_sge.num_sge)
 406                                 *sge = *sqp->s_sge.sg_list++;
 407                 } else if (sge->length == 0 && sge->mr != NULL) {
 408                         if (++sge->n >= IPATH_SEGSZ) {
 409                                 if (++sge->m >= sge->mr->mapsz)
 410                                         break;
 411                                 sge->n = 0;
 412                         }
 413                         sge->vaddr =
 414                                 sge->mr->map[sge->m]->segs[sge->n].vaddr;
 415                         sge->length =
 416                                 sge->mr->map[sge->m]->segs[sge->n].length;
 417                 }
 418                 sqp->s_len -= len;
 419         }
 420
 421         if (wqe->wr.opcode == IB_WR_RDMA_WRITE ||
 422             wqe->wr.opcode == IB_WR_RDMA_READ)
 423                 goto send_comp;
 424
 425         if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM)
 426                 wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
 427         else
 428                 wc.opcode = IB_WC_RECV;
 429         wc.wr_id = qp->r_wr_id;
 430         wc.status = IB_WC_SUCCESS;
 431         wc.vendor_err = 0;
 432         wc.byte_len = wqe->length;
 433         wc.qp = &qp->ibqp;
 434         wc.src_qp = qp->remote_qpn;
 435         /* XXX do we know which pkey matched? Only needed for GSI. */
 436         wc.pkey_index = 0;
 437         wc.slid = qp->remote_ah_attr.dlid;
 438         wc.sl = qp->remote_ah_attr.sl;
 439         wc.dlid_path_bits = 0;
 440         /* Signal completion event if the solicited bit is set. */
 441         ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
 442                        wqe->wr.send_flags & IB_SEND_SOLICITED);
 443
 444 send_comp:
 445         sqp->s_rnr_retry = sqp->s_rnr_retry_cnt;
 446
 447         if (!(sqp->s_flags & IPATH_S_SIGNAL_REQ_WR) ||
 448             (wqe->wr.send_flags & IB_SEND_SIGNALED)) {
 449                 wc.wr_id = wqe->wr.wr_id;
 450                 wc.status = IB_WC_SUCCESS;
 451                 wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
 452                 wc.vendor_err = 0;
 453                 wc.byte_len = wqe->length;
 454                 wc.qp = &sqp->ibqp;
 455                 wc.src_qp = 0;
 456                 wc.pkey_index = 0;
 457                 wc.slid = 0;
 458                 wc.sl = 0;
 459                 wc.dlid_path_bits = 0;
 460                 wc.port_num = 0;
 461                 ipath_cq_enter(to_icq(sqp->ibqp.send_cq), &wc, 0);
 462         }
 463
 464         /* Update s_last now that we are finished with the SWQE */
 465         spin_lock_irqsave(&sqp->s_lock, flags);
 466         if (++sqp->s_last >= sqp->s_size)
 467                 sqp->s_last = 0;
 468         spin_unlock_irqrestore(&sqp->s_lock, flags);
 469         goto again;
 470
 471 done:
 472         if (atomic_dec_and_test(&qp->refcount))
 473                 wake_up(&qp->wait);
 474 }
 475
 476 static int want_buffer(struct ipath_devdata *dd)
 477 {
 478         set_bit(IPATH_S_PIOINTBUFAVAIL, &dd->ipath_sendctrl);
 479         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
 480                          dd->ipath_sendctrl);
 481
 482         return 0;
 483 }
 484
 485 /**
 486  * ipath_no_bufs_available - tell the layer driver we need buffers
 487  * @qp: the QP that caused the problem
 488  * @dev: the device we ran out of buffers on
 489  *
 490  * Called when we run out of PIO buffers.
 491  */
 492 void ipath_no_bufs_available(struct ipath_qp *qp, struct ipath_ibdev *dev)
 493 {
 494         unsigned long flags;
 495
 496         spin_lock_irqsave(&dev->pending_lock, flags);
 497         if (list_empty(&qp->piowait))
 498                 list_add_tail(&qp->piowait, &dev->piowait);
 499         spin_unlock_irqrestore(&dev->pending_lock, flags);
 500         /*
 501          * Note that as soon as want_buffer() is called and
 502          * possibly before it returns, ipath_ib_piobufavail()
 503          * could be called.  If we are still in the tasklet function,
 504          * tasklet_hi_schedule() will not call us until the next time
 505          * tasklet_hi_schedule() is called.
 506          * We clear the tasklet flag now since we are committing to return
 507          * from the tasklet function.
 508          */
 509         clear_bit(IPATH_S_BUSY, &qp->s_busy);
 510         tasklet_unlock(&qp->s_task);
 511         want_buffer(dev->dd);
 512         dev->n_piowait++;
 513 }
 514
 515 /**
 516  * ipath_post_ruc_send - post RC and UC sends
 517  * @qp: the QP to post on
 518  * @wr: the work request to send
 519  */
 520 int ipath_post_ruc_send(struct ipath_qp *qp, struct ib_send_wr *wr)
 521 {
 522         struct ipath_swqe *wqe;
 523         unsigned long flags;
 524         u32 next;
 525         int i, j;
 526         int acc;
 527         int ret;
 528
 529         /*
 530          * Don't allow RDMA reads or atomic operations on UC or
 531          * undefined operations.
 532          * Make sure buffer is large enough to hold the result for atomics.
 533          */
 534         if (qp->ibqp.qp_type == IB_QPT_UC) {
 535                 if ((unsigned) wr->opcode >= IB_WR_RDMA_READ) {
 536                         ret = -EINVAL;
 537                         goto bail;
 538                 }
 539         } else if ((unsigned) wr->opcode > IB_WR_ATOMIC_FETCH_AND_ADD) {
 540                 ret = -EINVAL;
 541                 goto bail;
 542         } else if (wr->opcode >= IB_WR_ATOMIC_CMP_AND_SWP &&
 543                    (wr->num_sge == 0 ||
 544                     wr->sg_list[0].length < sizeof(u64) ||
 545                     wr->sg_list[0].addr & (sizeof(u64) - 1))) {
 546                 ret = -EINVAL;
 547                 goto bail;
 548         } else if (wr->opcode >= IB_WR_RDMA_READ && !qp->s_max_rd_atomic) {
 549                 ret = -EINVAL;
 550                 goto bail;
 551         }
 552         /* IB spec says that num_sge == 0 is OK. */
 553         if (wr->num_sge > qp->s_max_sge) {
 554                 ret = -ENOMEM;
 555                 goto bail;
 556         }
 557         spin_lock_irqsave(&qp->s_lock, flags);
 558         next = qp->s_head + 1;
 559         if (next >= qp->s_size)
 560                 next = 0;
 561         if (next == qp->s_last) {
 562                 spin_unlock_irqrestore(&qp->s_lock, flags);
 563                 ret = -EINVAL;
 564                 goto bail;
 565         }
 566
 567         wqe = get_swqe_ptr(qp, qp->s_head);
 568         wqe->wr = *wr;
 569         wqe->ssn = qp->s_ssn++;
 570         wqe->sg_list[0].mr = NULL;
 571         wqe->sg_list[0].vaddr = NULL;
 572         wqe->sg_list[0].length = 0;
 573         wqe->sg_list[0].sge_length = 0;
 574         wqe->length = 0;
 575         acc = wr->opcode >= IB_WR_RDMA_READ ? IB_ACCESS_LOCAL_WRITE : 0;
 576         for (i = 0, j = 0; i < wr->num_sge; i++) {
 577                 if (to_ipd(qp->ibqp.pd)->user && wr->sg_list[i].lkey == 0) {
 578                         spin_unlock_irqrestore(&qp->s_lock, flags);
 579                         ret = -EINVAL;
 580                         goto bail;
 581                 }
 582                 if (wr->sg_list[i].length == 0)
 583                         continue;
 584                 if (!ipath_lkey_ok(qp, &wqe->sg_list[j], &wr->sg_list[i],
 585                                    acc)) {
 586                         spin_unlock_irqrestore(&qp->s_lock, flags);
 587                         ret = -EINVAL;
 588                         goto bail;
 589                 }
 590                 wqe->length += wr->sg_list[i].length;
 591                 j++;
 592         }
 593         wqe->wr.num_sge = j;
 594         qp->s_head = next;
 595         spin_unlock_irqrestore(&qp->s_lock, flags);
 596
 597         ipath_do_ruc_send((unsigned long) qp);
 598
 599         ret = 0;
 600
 601 bail:
 602         return ret;
 603 }
 604
 605 /**
 606  * ipath_make_grh - construct a GRH header
 607  * @dev: a pointer to the ipath device
 608  * @hdr: a pointer to the GRH header being constructed
 609  * @grh: the global route address to send to
 610  * @hwords: the number of 32 bit words of header being sent
 611  * @nwords: the number of 32 bit words of data being sent
 612  *
 613  * Return the size of the header in 32 bit words.
 614  */
 615 u32 ipath_make_grh(struct ipath_ibdev *dev, struct ib_grh *hdr,
 616                    struct ib_global_route *grh, u32 hwords, u32 nwords)
 617 {
 618         hdr->version_tclass_flow =
 619                 cpu_to_be32((6 << 28) |
 620                             (grh->traffic_class << 20) |
 621                             grh->flow_label);
 622         hdr->paylen = cpu_to_be16((hwords - 2 + nwords + SIZE_OF_CRC) << 2);
 623         /* next_hdr is defined by C8-7 in ch. 8.4.1 */
 624         hdr->next_hdr = 0x1B;
 625         hdr->hop_limit = grh->hop_limit;
 626         /* The SGID is 32-bit aligned. */
 627         hdr->sgid.global.subnet_prefix = dev->gid_prefix;
 628         hdr->sgid.global.interface_id = dev->dd->ipath_guid;
 629         hdr->dgid = grh->dgid;
 630
 631         /* GRH header size in 32-bit words. */
 632         return sizeof(struct ib_grh) / sizeof(u32);
 633 }
 634
 635 /**
 636  * ipath_do_ruc_send - perform a send on an RC or UC QP
 637  * @data: contains a pointer to the QP
 638  *
 639  * Process entries in the send work queue until credit or queue is
 640  * exhausted.  Only allow one CPU to send a packet per QP (tasklet).
 641  * Otherwise, after we drop the QP s_lock, two threads could send
 642  * packets out of order.
 643  */
 644 void ipath_do_ruc_send(unsigned long data)
 645 {
 646         struct ipath_qp *qp = (struct ipath_qp *)data;
 647         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
 648         unsigned long flags;
 649         u16 lrh0;
 650         u32 nwords;
 651         u32 extra_bytes;
 652         u32 bth0;
 653         u32 bth2;
 654         u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
 655         struct ipath_other_headers *ohdr;
 656
 657         if (test_and_set_bit(IPATH_S_BUSY, &qp->s_busy))
 658                 goto bail;
 659
 660         if (unlikely(qp->remote_ah_attr.dlid == dev->dd->ipath_lid)) {
 661                 ipath_ruc_loopback(qp);
 662                 goto clear;
 663         }
 664
 665         ohdr = &qp->s_hdr.u.oth;
 666         if (qp->remote_ah_attr.ah_flags & IB_AH_GRH)
 667                 ohdr = &qp->s_hdr.u.l.oth;
 668
 669 again:
 670         /* Check for a constructed packet to be sent. */
 671         if (qp->s_hdrwords != 0) {
 672                 /*
 673                  * If no PIO bufs are available, return.  An interrupt will
 674                  * call ipath_ib_piobufavail() when one is available.
 675                  */
 676                 if (ipath_verbs_send(dev->dd, qp->s_hdrwords,
 677                                      (u32 *) &qp->s_hdr, qp->s_cur_size,
 678                                      qp->s_cur_sge)) {
 679                         ipath_no_bufs_available(qp, dev);
 680                         goto bail;
 681                 }
 682                 dev->n_unicast_xmit++;
 683                 /* Record that we sent the packet and s_hdr is empty. */
 684                 qp->s_hdrwords = 0;
 685         }
 686
 687         /*
 688          * The lock is needed to synchronize between setting
 689          * qp->s_ack_state, resend timer, and post_send().
 690          */
 691         spin_lock_irqsave(&qp->s_lock, flags);
 692
 693         if (!((qp->ibqp.qp_type == IB_QPT_RC) ?
 694                ipath_make_rc_req(qp, ohdr, pmtu, &bth0, &bth2) :
 695                ipath_make_uc_req(qp, ohdr, pmtu, &bth0, &bth2))) {
 696                 /*
 697                  * Clear the busy bit before unlocking to avoid races with
 698                  * adding new work queue items and then failing to process
 699                  * them.
 700                  */
 701                 clear_bit(IPATH_S_BUSY, &qp->s_busy);
 702                 spin_unlock_irqrestore(&qp->s_lock, flags);
 703                 goto bail;
 704         }
 705
 706         spin_unlock_irqrestore(&qp->s_lock, flags);
 707
 708         /* Construct the header. */
 709         extra_bytes = (4 - qp->s_cur_size) & 3;
 710         nwords = (qp->s_cur_size + extra_bytes) >> 2;
 711         lrh0 = IPATH_LRH_BTH;
 712         if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
 713                 qp->s_hdrwords += ipath_make_grh(dev, &qp->s_hdr.u.l.grh,
 714                                                  &qp->remote_ah_attr.grh,
 715                                                  qp->s_hdrwords, nwords);
 716                 lrh0 = IPATH_LRH_GRH;
 717         }
 718         lrh0 |= qp->remote_ah_attr.sl << 4;
 719         qp->s_hdr.lrh[0] = cpu_to_be16(lrh0);
 720         qp->s_hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
 721         qp->s_hdr.lrh[2] = cpu_to_be16(qp->s_hdrwords + nwords +
 722                                        SIZE_OF_CRC);
 723         qp->s_hdr.lrh[3] = cpu_to_be16(dev->dd->ipath_lid);
 724         bth0 |= ipath_get_pkey(dev->dd, qp->s_pkey_index);
 725         bth0 |= extra_bytes << 20;
 726         ohdr->bth[0] = cpu_to_be32(bth0);
 727         ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
 728         ohdr->bth[2] = cpu_to_be32(bth2);
 729
 730         /* Check for more work to do. */
 731         goto again;
 732
 733 clear:
 734         clear_bit(IPATH_S_BUSY, &qp->s_busy);
 735 bail:
 736         return;
 737 }