err.no 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 "ips_common.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 /**
 110  * ipath_get_rwqe - copy the next RWQE into the QP's RWQE
 111  * @qp: the QP
 112  * @wr_id_only: update wr_id only, not SGEs
 113  *
 114  * Return 0 if no RWQE is available, otherwise return 1.
 115  *
 116  * Called at interrupt level with the QP r_rq.lock held.
 117  */
 118 int ipath_get_rwqe(struct ipath_qp *qp, int wr_id_only)
 119 {
 120         struct ipath_rq *rq;
 121         struct ipath_srq *srq;
 122         struct ipath_rwqe *wqe;
 123         int ret;
 124
 125         if (!qp->ibqp.srq) {
 126                 rq = &qp->r_rq;
 127                 if (unlikely(rq->tail == rq->head)) {
 128                         ret = 0;
 129                         goto bail;
 130                 }
 131                 wqe = get_rwqe_ptr(rq, rq->tail);
 132                 qp->r_wr_id = wqe->wr_id;
 133                 if (!wr_id_only) {
 134                         qp->r_sge.sge = wqe->sg_list[0];
 135                         qp->r_sge.sg_list = wqe->sg_list + 1;
 136                         qp->r_sge.num_sge = wqe->num_sge;
 137                         qp->r_len = wqe->length;
 138                 }
 139                 if (++rq->tail >= rq->size)
 140                         rq->tail = 0;
 141                 ret = 1;
 142                 goto bail;
 143         }
 144
 145         srq = to_isrq(qp->ibqp.srq);
 146         rq = &srq->rq;
 147         spin_lock(&rq->lock);
 148         if (unlikely(rq->tail == rq->head)) {
 149                 spin_unlock(&rq->lock);
 150                 ret = 0;
 151                 goto bail;
 152         }
 153         wqe = get_rwqe_ptr(rq, rq->tail);
 154         qp->r_wr_id = wqe->wr_id;
 155         if (!wr_id_only) {
 156                 qp->r_sge.sge = wqe->sg_list[0];
 157                 qp->r_sge.sg_list = wqe->sg_list + 1;
 158                 qp->r_sge.num_sge = wqe->num_sge;
 159                 qp->r_len = wqe->length;
 160         }
 161         if (++rq->tail >= rq->size)
 162                 rq->tail = 0;
 163         if (srq->ibsrq.event_handler) {
 164                 struct ib_event ev;
 165                 u32 n;
 166
 167                 if (rq->head < rq->tail)
 168                         n = rq->size + rq->head - rq->tail;
 169                 else
 170                         n = rq->head - rq->tail;
 171                 if (n < srq->limit) {
 172                         srq->limit = 0;
 173                         spin_unlock(&rq->lock);
 174                         ev.device = qp->ibqp.device;
 175                         ev.element.srq = qp->ibqp.srq;
 176                         ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
 177                         srq->ibsrq.event_handler(&ev,
 178                                                  srq->ibsrq.srq_context);
 179                 } else
 180                         spin_unlock(&rq->lock);
 181         } else
 182                 spin_unlock(&rq->lock);
 183         ret = 1;
 184
 185 bail:
 186         return ret;
 187 }
 188
 189 /**
 190  * ipath_ruc_loopback - handle UC and RC lookback requests
 191  * @sqp: the loopback QP
 192  *
 193  * This is called from ipath_do_uc_send() or ipath_do_rc_send() to
 194  * forward a WQE addressed to the same HCA.
 195  * Note that although we are single threaded due to the tasklet, we still
 196  * have to protect against post_send().  We don't have to worry about
 197  * receive interrupts since this is a connected protocol and all packets
 198  * will pass through here.
 199  */
 200 static void ipath_ruc_loopback(struct ipath_qp *sqp)
 201 {
 202         struct ipath_ibdev *dev = to_idev(sqp->ibqp.device);
 203         struct ipath_qp *qp;
 204         struct ipath_swqe *wqe;
 205         struct ipath_sge *sge;
 206         unsigned long flags;
 207         struct ib_wc wc;
 208         u64 sdata;
 209
 210         qp = ipath_lookup_qpn(&dev->qp_table, sqp->remote_qpn);
 211         if (!qp) {
 212                 dev->n_pkt_drops++;
 213                 return;
 214         }
 215
 216 again:
 217         spin_lock_irqsave(&sqp->s_lock, flags);
 218
 219         if (!(ib_ipath_state_ops[sqp->state] & IPATH_PROCESS_SEND_OK)) {
 220                 spin_unlock_irqrestore(&sqp->s_lock, flags);
 221                 goto done;
 222         }
 223
 224         /* Get the next send request. */
 225         if (sqp->s_last == sqp->s_head) {
 226                 /* Send work queue is empty. */
 227                 spin_unlock_irqrestore(&sqp->s_lock, flags);
 228                 goto done;
 229         }
 230
 231         /*
 232          * We can rely on the entry not changing without the s_lock
 233          * being held until we update s_last.
 234          */
 235         wqe = get_swqe_ptr(sqp, sqp->s_last);
 236         spin_unlock_irqrestore(&sqp->s_lock, flags);
 237
 238         wc.wc_flags = 0;
 239         wc.imm_data = 0;
 240
 241         sqp->s_sge.sge = wqe->sg_list[0];
 242         sqp->s_sge.sg_list = wqe->sg_list + 1;
 243         sqp->s_sge.num_sge = wqe->wr.num_sge;
 244         sqp->s_len = wqe->length;
 245         switch (wqe->wr.opcode) {
 246         case IB_WR_SEND_WITH_IMM:
 247                 wc.wc_flags = IB_WC_WITH_IMM;
 248                 wc.imm_data = wqe->wr.imm_data;
 249                 /* FALLTHROUGH */
 250         case IB_WR_SEND:
 251                 spin_lock_irqsave(&qp->r_rq.lock, flags);
 252                 if (!ipath_get_rwqe(qp, 0)) {
 253                 rnr_nak:
 254                         spin_unlock_irqrestore(&qp->r_rq.lock, flags);
 255                         /* Handle RNR NAK */
 256                         if (qp->ibqp.qp_type == IB_QPT_UC)
 257                                 goto send_comp;
 258                         if (sqp->s_rnr_retry == 0) {
 259                                 wc.status = IB_WC_RNR_RETRY_EXC_ERR;
 260                                 goto err;
 261                         }
 262                         if (sqp->s_rnr_retry_cnt < 7)
 263                                 sqp->s_rnr_retry--;
 264                         dev->n_rnr_naks++;
 265                         sqp->s_rnr_timeout =
 266                                 ib_ipath_rnr_table[sqp->s_min_rnr_timer];
 267                         ipath_insert_rnr_queue(sqp);
 268                         goto done;
 269                 }
 270                 spin_unlock_irqrestore(&qp->r_rq.lock, flags);
 271                 break;
 272
 273         case IB_WR_RDMA_WRITE_WITH_IMM:
 274                 wc.wc_flags = IB_WC_WITH_IMM;
 275                 wc.imm_data = wqe->wr.imm_data;
 276                 spin_lock_irqsave(&qp->r_rq.lock, flags);
 277                 if (!ipath_get_rwqe(qp, 1))
 278                         goto rnr_nak;
 279                 spin_unlock_irqrestore(&qp->r_rq.lock, flags);
 280                 /* FALLTHROUGH */
 281         case IB_WR_RDMA_WRITE:
 282                 if (wqe->length == 0)
 283                         break;
 284                 if (unlikely(!ipath_rkey_ok(dev, &qp->r_sge, wqe->length,
 285                                             wqe->wr.wr.rdma.remote_addr,
 286                                             wqe->wr.wr.rdma.rkey,
 287                                             IB_ACCESS_REMOTE_WRITE))) {
 288                 acc_err:
 289                         wc.status = IB_WC_REM_ACCESS_ERR;
 290                 err:
 291                         wc.wr_id = wqe->wr.wr_id;
 292                         wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
 293                         wc.vendor_err = 0;
 294                         wc.byte_len = 0;
 295                         wc.qp_num = sqp->ibqp.qp_num;
 296                         wc.src_qp = sqp->remote_qpn;
 297                         wc.pkey_index = 0;
 298                         wc.slid = sqp->remote_ah_attr.dlid;
 299                         wc.sl = sqp->remote_ah_attr.sl;
 300                         wc.dlid_path_bits = 0;
 301                         wc.port_num = 0;
 302                         ipath_sqerror_qp(sqp, &wc);
 303                         goto done;
 304                 }
 305                 break;
 306
 307         case IB_WR_RDMA_READ:
 308                 if (unlikely(!ipath_rkey_ok(dev, &sqp->s_sge, wqe->length,
 309                                             wqe->wr.wr.rdma.remote_addr,
 310                                             wqe->wr.wr.rdma.rkey,
 311                                             IB_ACCESS_REMOTE_READ)))
 312                         goto acc_err;
 313                 if (unlikely(!(qp->qp_access_flags &
 314                                IB_ACCESS_REMOTE_READ)))
 315                         goto acc_err;
 316                 qp->r_sge.sge = wqe->sg_list[0];
 317                 qp->r_sge.sg_list = wqe->sg_list + 1;
 318                 qp->r_sge.num_sge = wqe->wr.num_sge;
 319                 break;
 320
 321         case IB_WR_ATOMIC_CMP_AND_SWP:
 322         case IB_WR_ATOMIC_FETCH_AND_ADD:
 323                 if (unlikely(!ipath_rkey_ok(dev, &qp->r_sge, sizeof(u64),
 324                                             wqe->wr.wr.rdma.remote_addr,
 325                                             wqe->wr.wr.rdma.rkey,
 326                                             IB_ACCESS_REMOTE_ATOMIC)))
 327                         goto acc_err;
 328                 /* Perform atomic OP and save result. */
 329                 sdata = wqe->wr.wr.atomic.swap;
 330                 spin_lock_irqsave(&dev->pending_lock, flags);
 331                 qp->r_atomic_data = *(u64 *) qp->r_sge.sge.vaddr;
 332                 if (wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)
 333                         *(u64 *) qp->r_sge.sge.vaddr =
 334                                 qp->r_atomic_data + sdata;
 335                 else if (qp->r_atomic_data == wqe->wr.wr.atomic.compare_add)
 336                         *(u64 *) qp->r_sge.sge.vaddr = sdata;
 337                 spin_unlock_irqrestore(&dev->pending_lock, flags);
 338                 *(u64 *) sqp->s_sge.sge.vaddr = qp->r_atomic_data;
 339                 goto send_comp;
 340
 341         default:
 342                 goto done;
 343         }
 344
 345         sge = &sqp->s_sge.sge;
 346         while (sqp->s_len) {
 347                 u32 len = sqp->s_len;
 348
 349                 if (len > sge->length)
 350                         len = sge->length;
 351                 BUG_ON(len == 0);
 352                 ipath_copy_sge(&qp->r_sge, sge->vaddr, len);
 353                 sge->vaddr += len;
 354                 sge->length -= len;
 355                 sge->sge_length -= len;
 356                 if (sge->sge_length == 0) {
 357                         if (--sqp->s_sge.num_sge)
 358                                 *sge = *sqp->s_sge.sg_list++;
 359                 } else if (sge->length == 0 && sge->mr != NULL) {
 360                         if (++sge->n >= IPATH_SEGSZ) {
 361                                 if (++sge->m >= sge->mr->mapsz)
 362                                         break;
 363                                 sge->n = 0;
 364                         }
 365                         sge->vaddr =
 366                                 sge->mr->map[sge->m]->segs[sge->n].vaddr;
 367                         sge->length =
 368                                 sge->mr->map[sge->m]->segs[sge->n].length;
 369                 }
 370                 sqp->s_len -= len;
 371         }
 372
 373         if (wqe->wr.opcode == IB_WR_RDMA_WRITE ||
 374             wqe->wr.opcode == IB_WR_RDMA_READ)
 375                 goto send_comp;
 376
 377         if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM)
 378                 wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
 379         else
 380                 wc.opcode = IB_WC_RECV;
 381         wc.wr_id = qp->r_wr_id;
 382         wc.status = IB_WC_SUCCESS;
 383         wc.vendor_err = 0;
 384         wc.byte_len = wqe->length;
 385         wc.qp_num = qp->ibqp.qp_num;
 386         wc.src_qp = qp->remote_qpn;
 387         /* XXX do we know which pkey matched? Only needed for GSI. */
 388         wc.pkey_index = 0;
 389         wc.slid = qp->remote_ah_attr.dlid;
 390         wc.sl = qp->remote_ah_attr.sl;
 391         wc.dlid_path_bits = 0;
 392         /* Signal completion event if the solicited bit is set. */
 393         ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
 394                        wqe->wr.send_flags & IB_SEND_SOLICITED);
 395
 396 send_comp:
 397         sqp->s_rnr_retry = sqp->s_rnr_retry_cnt;
 398
 399         if (!test_bit(IPATH_S_SIGNAL_REQ_WR, &sqp->s_flags) ||
 400             (wqe->wr.send_flags & IB_SEND_SIGNALED)) {
 401                 wc.wr_id = wqe->wr.wr_id;
 402                 wc.status = IB_WC_SUCCESS;
 403                 wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
 404                 wc.vendor_err = 0;
 405                 wc.byte_len = wqe->length;
 406                 wc.qp_num = sqp->ibqp.qp_num;
 407                 wc.src_qp = 0;
 408                 wc.pkey_index = 0;
 409                 wc.slid = 0;
 410                 wc.sl = 0;
 411                 wc.dlid_path_bits = 0;
 412                 wc.port_num = 0;
 413                 ipath_cq_enter(to_icq(sqp->ibqp.send_cq), &wc, 0);
 414         }
 415
 416         /* Update s_last now that we are finished with the SWQE */
 417         spin_lock_irqsave(&sqp->s_lock, flags);
 418         if (++sqp->s_last >= sqp->s_size)
 419                 sqp->s_last = 0;
 420         spin_unlock_irqrestore(&sqp->s_lock, flags);
 421         goto again;
 422
 423 done:
 424         if (atomic_dec_and_test(&qp->refcount))
 425                 wake_up(&qp->wait);
 426 }
 427
 428 /**
 429  * ipath_no_bufs_available - tell the layer driver we need buffers
 430  * @qp: the QP that caused the problem
 431  * @dev: the device we ran out of buffers on
 432  *
 433  * Called when we run out of PIO buffers.
 434  */
 435 void ipath_no_bufs_available(struct ipath_qp *qp, struct ipath_ibdev *dev)
 436 {
 437         unsigned long flags;
 438
 439         spin_lock_irqsave(&dev->pending_lock, flags);
 440         if (list_empty(&qp->piowait))
 441                 list_add_tail(&qp->piowait, &dev->piowait);
 442         spin_unlock_irqrestore(&dev->pending_lock, flags);
 443         /*
 444          * Note that as soon as ipath_layer_want_buffer() is called and
 445          * possibly before it returns, ipath_ib_piobufavail()
 446          * could be called.  If we are still in the tasklet function,
 447          * tasklet_hi_schedule() will not call us until the next time
 448          * tasklet_hi_schedule() is called.
 449          * We clear the tasklet flag now since we are committing to return
 450          * from the tasklet function.
 451          */
 452         clear_bit(IPATH_S_BUSY, &qp->s_flags);
 453         tasklet_unlock(&qp->s_task);
 454         ipath_layer_want_buffer(dev->dd);
 455         dev->n_piowait++;
 456 }
 457
 458 /**
 459  * ipath_post_ruc_send - post RC and UC sends
 460  * @qp: the QP to post on
 461  * @wr: the work request to send
 462  */
 463 int ipath_post_ruc_send(struct ipath_qp *qp, struct ib_send_wr *wr)
 464 {
 465         struct ipath_swqe *wqe;
 466         unsigned long flags;
 467         u32 next;
 468         int i, j;
 469         int acc;
 470         int ret;
 471
 472         /*
 473          * Don't allow RDMA reads or atomic operations on UC or
 474          * undefined operations.
 475          * Make sure buffer is large enough to hold the result for atomics.
 476          */
 477         if (qp->ibqp.qp_type == IB_QPT_UC) {
 478                 if ((unsigned) wr->opcode >= IB_WR_RDMA_READ) {
 479                         ret = -EINVAL;
 480                         goto bail;
 481                 }
 482         } else if ((unsigned) wr->opcode > IB_WR_ATOMIC_FETCH_AND_ADD) {
 483                 ret = -EINVAL;
 484                 goto bail;
 485         } else if (wr->opcode >= IB_WR_ATOMIC_CMP_AND_SWP &&
 486                    (wr->num_sge == 0 ||
 487                     wr->sg_list[0].length < sizeof(u64) ||
 488                     wr->sg_list[0].addr & (sizeof(u64) - 1))) {
 489                 ret = -EINVAL;
 490                 goto bail;
 491         }
 492         /* IB spec says that num_sge == 0 is OK. */
 493         if (wr->num_sge > qp->s_max_sge) {
 494                 ret = -ENOMEM;
 495                 goto bail;
 496         }
 497         spin_lock_irqsave(&qp->s_lock, flags);
 498         next = qp->s_head + 1;
 499         if (next >= qp->s_size)
 500                 next = 0;
 501         if (next == qp->s_last) {
 502                 spin_unlock_irqrestore(&qp->s_lock, flags);
 503                 ret = -EINVAL;
 504                 goto bail;
 505         }
 506
 507         wqe = get_swqe_ptr(qp, qp->s_head);
 508         wqe->wr = *wr;
 509         wqe->ssn = qp->s_ssn++;
 510         wqe->sg_list[0].mr = NULL;
 511         wqe->sg_list[0].vaddr = NULL;
 512         wqe->sg_list[0].length = 0;
 513         wqe->sg_list[0].sge_length = 0;
 514         wqe->length = 0;
 515         acc = wr->opcode >= IB_WR_RDMA_READ ? IB_ACCESS_LOCAL_WRITE : 0;
 516         for (i = 0, j = 0; i < wr->num_sge; i++) {
 517                 if (to_ipd(qp->ibqp.pd)->user && wr->sg_list[i].lkey == 0) {
 518                         spin_unlock_irqrestore(&qp->s_lock, flags);
 519                         ret = -EINVAL;
 520                         goto bail;
 521                 }
 522                 if (wr->sg_list[i].length == 0)
 523                         continue;
 524                 if (!ipath_lkey_ok(&to_idev(qp->ibqp.device)->lk_table,
 525                                    &wqe->sg_list[j], &wr->sg_list[i],
 526                                    acc)) {
 527                         spin_unlock_irqrestore(&qp->s_lock, flags);
 528                         ret = -EINVAL;
 529                         goto bail;
 530                 }
 531                 wqe->length += wr->sg_list[i].length;
 532                 j++;
 533         }
 534         wqe->wr.num_sge = j;
 535         qp->s_head = next;
 536         spin_unlock_irqrestore(&qp->s_lock, flags);
 537
 538         ipath_do_ruc_send((unsigned long) qp);
 539
 540         ret = 0;
 541
 542 bail:
 543         return ret;
 544 }
 545
 546 /**
 547  * ipath_make_grh - construct a GRH header
 548  * @dev: a pointer to the ipath device
 549  * @hdr: a pointer to the GRH header being constructed
 550  * @grh: the global route address to send to
 551  * @hwords: the number of 32 bit words of header being sent
 552  * @nwords: the number of 32 bit words of data being sent
 553  *
 554  * Return the size of the header in 32 bit words.
 555  */
 556 u32 ipath_make_grh(struct ipath_ibdev *dev, struct ib_grh *hdr,
 557                    struct ib_global_route *grh, u32 hwords, u32 nwords)
 558 {
 559         hdr->version_tclass_flow =
 560                 cpu_to_be32((6 << 28) |
 561                             (grh->traffic_class << 20) |
 562                             grh->flow_label);
 563         hdr->paylen = cpu_to_be16((hwords - 2 + nwords + SIZE_OF_CRC) << 2);
 564         /* next_hdr is defined by C8-7 in ch. 8.4.1 */
 565         hdr->next_hdr = 0x1B;
 566         hdr->hop_limit = grh->hop_limit;
 567         /* The SGID is 32-bit aligned. */
 568         hdr->sgid.global.subnet_prefix = dev->gid_prefix;
 569         hdr->sgid.global.interface_id = ipath_layer_get_guid(dev->dd);
 570         hdr->dgid = grh->dgid;
 571
 572         /* GRH header size in 32-bit words. */
 573         return sizeof(struct ib_grh) / sizeof(u32);
 574 }
 575
 576 /**
 577  * ipath_do_ruc_send - perform a send on an RC or UC QP
 578  * @data: contains a pointer to the QP
 579  *
 580  * Process entries in the send work queue until credit or queue is
 581  * exhausted.  Only allow one CPU to send a packet per QP (tasklet).
 582  * Otherwise, after we drop the QP s_lock, two threads could send
 583  * packets out of order.
 584  */
 585 void ipath_do_ruc_send(unsigned long data)
 586 {
 587         struct ipath_qp *qp = (struct ipath_qp *)data;
 588         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
 589         unsigned long flags;
 590         u16 lrh0;
 591         u32 nwords;
 592         u32 extra_bytes;
 593         u32 bth0;
 594         u32 bth2;
 595         u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
 596         struct ipath_other_headers *ohdr;
 597
 598         if (test_and_set_bit(IPATH_S_BUSY, &qp->s_flags))
 599                 goto bail;
 600
 601         if (unlikely(qp->remote_ah_attr.dlid ==
 602                      ipath_layer_get_lid(dev->dd))) {
 603                 ipath_ruc_loopback(qp);
 604                 goto clear;
 605         }
 606
 607         ohdr = &qp->s_hdr.u.oth;
 608         if (qp->remote_ah_attr.ah_flags & IB_AH_GRH)
 609                 ohdr = &qp->s_hdr.u.l.oth;
 610
 611 again:
 612         /* Check for a constructed packet to be sent. */
 613         if (qp->s_hdrwords != 0) {
 614                 /*
 615                  * If no PIO bufs are available, return.  An interrupt will
 616                  * call ipath_ib_piobufavail() when one is available.
 617                  */
 618                 if (ipath_verbs_send(dev->dd, qp->s_hdrwords,
 619                                      (u32 *) &qp->s_hdr, qp->s_cur_size,
 620                                      qp->s_cur_sge)) {
 621                         ipath_no_bufs_available(qp, dev);
 622                         goto bail;
 623                 }
 624                 dev->n_unicast_xmit++;
 625                 /* Record that we sent the packet and s_hdr is empty. */
 626                 qp->s_hdrwords = 0;
 627         }
 628
 629         /*
 630          * The lock is needed to synchronize between setting
 631          * qp->s_ack_state, resend timer, and post_send().
 632          */
 633         spin_lock_irqsave(&qp->s_lock, flags);
 634
 635         /* Sending responses has higher priority over sending requests. */
 636         if (qp->s_ack_state != IB_OPCODE_RC_ACKNOWLEDGE &&
 637             (bth0 = ipath_make_rc_ack(qp, ohdr, pmtu)) != 0)
 638                 bth2 = qp->s_ack_psn++ & IPS_PSN_MASK;
 639         else if (!((qp->ibqp.qp_type == IB_QPT_RC) ?
 640                    ipath_make_rc_req(qp, ohdr, pmtu, &bth0, &bth2) :
 641                    ipath_make_uc_req(qp, ohdr, pmtu, &bth0, &bth2))) {
 642                 /*
 643                  * Clear the busy bit before unlocking to avoid races with
 644                  * adding new work queue items and then failing to process
 645                  * them.
 646                  */
 647                 clear_bit(IPATH_S_BUSY, &qp->s_flags);
 648                 spin_unlock_irqrestore(&qp->s_lock, flags);
 649                 goto bail;
 650         }
 651
 652         spin_unlock_irqrestore(&qp->s_lock, flags);
 653
 654         /* Construct the header. */
 655         extra_bytes = (4 - qp->s_cur_size) & 3;
 656         nwords = (qp->s_cur_size + extra_bytes) >> 2;
 657         lrh0 = IPS_LRH_BTH;
 658         if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
 659                 qp->s_hdrwords += ipath_make_grh(dev, &qp->s_hdr.u.l.grh,
 660                                                  &qp->remote_ah_attr.grh,
 661                                                  qp->s_hdrwords, nwords);
 662                 lrh0 = IPS_LRH_GRH;
 663         }
 664         lrh0 |= qp->remote_ah_attr.sl << 4;
 665         qp->s_hdr.lrh[0] = cpu_to_be16(lrh0);
 666         qp->s_hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
 667         qp->s_hdr.lrh[2] = cpu_to_be16(qp->s_hdrwords + nwords +
 668                                        SIZE_OF_CRC);
 669         qp->s_hdr.lrh[3] = cpu_to_be16(ipath_layer_get_lid(dev->dd));
 670         bth0 |= ipath_layer_get_pkey(dev->dd, qp->s_pkey_index);
 671         bth0 |= extra_bytes << 20;
 672         ohdr->bth[0] = cpu_to_be32(bth0);
 673         ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
 674         ohdr->bth[2] = cpu_to_be32(bth2);
 675
 676         /* Check for more work to do. */
 677         goto again;
 678
 679 clear:
 680         clear_bit(IPATH_S_BUSY, &qp->s_flags);
 681 bail:
 682         return;
 683 }