release/src-rt-6.x.4708/linux/linux-2.6.36/drivers/infiniband/hw/ipath/ipath_ruc.c

   1 /*
   2  * Copyright (c) 2006, 2007, 2008 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/sched.h>
  35 #include <linux/spinlock.h>
  36
  37 #include "ipath_verbs.h"
  38 #include "ipath_kernel.h"
  39
  40 /*
  41  * Convert the AETH RNR timeout code into the number of milliseconds.
  42  */
  43 const u32 ib_ipath_rnr_table[32] = {
  44         656,                    /* 0 */
  45         1,                      /* 1 */
  46         1,                      /* 2 */
  47         1,                      /* 3 */
  48         1,                      /* 4 */
  49         1,                      /* 5 */
  50         1,                      /* 6 */
  51         1,                      /* 7 */
  52         1,                      /* 8 */
  53         1,                      /* 9 */
  54         1,                      /* A */
  55         1,                      /* B */
  56         1,                      /* C */
  57         1,                      /* D */
  58         2,                      /* E */
  59         2,                      /* F */
  60         3,                      /* 10 */
  61         4,                      /* 11 */
  62         6,                      /* 12 */
  63         8,                      /* 13 */
  64         11,                     /* 14 */
  65         16,                     /* 15 */
  66         21,                     /* 16 */
  67         31,                     /* 17 */
  68         41,                     /* 18 */
  69         62,                     /* 19 */
  70         82,                     /* 1A */
  71         123,                    /* 1B */
  72         164,                    /* 1C */
  73         246,                    /* 1D */
  74         328,                    /* 1E */
  75         492                     /* 1F */
  76 };
  77
  78 void ipath_insert_rnr_queue(struct ipath_qp *qp)
  79 {
  80         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
  81
  82         /* We already did a spin_lock_irqsave(), so just use spin_lock */
  83         spin_lock(&dev->pending_lock);
  84         if (list_empty(&dev->rnrwait))
  85                 list_add(&qp->timerwait, &dev->rnrwait);
  86         else {
  87                 struct list_head *l = &dev->rnrwait;
  88                 struct ipath_qp *nqp = list_entry(l->next, struct ipath_qp,
  89                                                   timerwait);
  90
  91                 while (qp->s_rnr_timeout >= nqp->s_rnr_timeout) {
  92                         qp->s_rnr_timeout -= nqp->s_rnr_timeout;
  93                         l = l->next;
  94                         if (l->next == &dev->rnrwait) {
  95                                 nqp = NULL;
  96                                 break;
  97                         }
  98                         nqp = list_entry(l->next, struct ipath_qp,
  99                                          timerwait);
 100                 }
 101                 if (nqp)
 102                         nqp->s_rnr_timeout -= qp->s_rnr_timeout;
 103                 list_add(&qp->timerwait, l);
 104         }
 105         spin_unlock(&dev->pending_lock);
 106 }
 107
 108 /**
 109  * ipath_init_sge - Validate a RWQE and fill in the SGE state
 110  * @qp: the QP
 111  *
 112  * Return 1 if OK.
 113  */
 114 int ipath_init_sge(struct ipath_qp *qp, struct ipath_rwqe *wqe,
 115                    u32 *lengthp, struct ipath_sge_state *ss)
 116 {
 117         int i, j, ret;
 118         struct ib_wc wc;
 119
 120         *lengthp = 0;
 121         for (i = j = 0; i < wqe->num_sge; i++) {
 122                 if (wqe->sg_list[i].length == 0)
 123                         continue;
 124                 /* Check LKEY */
 125                 if (!ipath_lkey_ok(qp, j ? &ss->sg_list[j - 1] : &ss->sge,
 126                                    &wqe->sg_list[i], IB_ACCESS_LOCAL_WRITE))
 127                         goto bad_lkey;
 128                 *lengthp += wqe->sg_list[i].length;
 129                 j++;
 130         }
 131         ss->num_sge = j;
 132         ret = 1;
 133         goto bail;
 134
 135 bad_lkey:
 136         memset(&wc, 0, sizeof(wc));
 137         wc.wr_id = wqe->wr_id;
 138         wc.status = IB_WC_LOC_PROT_ERR;
 139         wc.opcode = IB_WC_RECV;
 140         wc.qp = &qp->ibqp;
 141         /* Signal solicited completion event. */
 142         ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
 143         ret = 0;
 144 bail:
 145         return ret;
 146 }
 147
 148 /**
 149  * ipath_get_rwqe - copy the next RWQE into the QP's RWQE
 150  * @qp: the QP
 151  * @wr_id_only: update qp->r_wr_id only, not qp->r_sge
 152  *
 153  * Return 0 if no RWQE is available, otherwise return 1.
 154  *
 155  * Can be called from interrupt level.
 156  */
 157 int ipath_get_rwqe(struct ipath_qp *qp, int wr_id_only)
 158 {
 159         unsigned long flags;
 160         struct ipath_rq *rq;
 161         struct ipath_rwq *wq;
 162         struct ipath_srq *srq;
 163         struct ipath_rwqe *wqe;
 164         void (*handler)(struct ib_event *, void *);
 165         u32 tail;
 166         int ret;
 167
 168         if (qp->ibqp.srq) {
 169                 srq = to_isrq(qp->ibqp.srq);
 170                 handler = srq->ibsrq.event_handler;
 171                 rq = &srq->rq;
 172         } else {
 173                 srq = NULL;
 174                 handler = NULL;
 175                 rq = &qp->r_rq;
 176         }
 177
 178         spin_lock_irqsave(&rq->lock, flags);
 179         if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_RECV_OK)) {
 180                 ret = 0;
 181                 goto unlock;
 182         }
 183
 184         wq = rq->wq;
 185         tail = wq->tail;
 186         /* Validate tail before using it since it is user writable. */
 187         if (tail >= rq->size)
 188                 tail = 0;
 189         do {
 190                 if (unlikely(tail == wq->head)) {
 191                         ret = 0;
 192                         goto unlock;
 193                 }
 194                 /* Make sure entry is read after head index is read. */
 195                 smp_rmb();
 196                 wqe = get_rwqe_ptr(rq, tail);
 197                 if (++tail >= rq->size)
 198                         tail = 0;
 199                 if (wr_id_only)
 200                         break;
 201                 qp->r_sge.sg_list = qp->r_sg_list;
 202         } while (!ipath_init_sge(qp, wqe, &qp->r_len, &qp->r_sge));
 203         qp->r_wr_id = wqe->wr_id;
 204         wq->tail = tail;
 205
 206         ret = 1;
 207         set_bit(IPATH_R_WRID_VALID, &qp->r_aflags);
 208         if (handler) {
 209                 u32 n;
 210
 211                 /*
 212                  * validate head pointer value and compute
 213                  * the number of remaining WQEs.
 214                  */
 215                 n = wq->head;
 216                 if (n >= rq->size)
 217                         n = 0;
 218                 if (n < tail)
 219                         n += rq->size - tail;
 220                 else
 221                         n -= tail;
 222                 if (n < srq->limit) {
 223                         struct ib_event ev;
 224
 225                         srq->limit = 0;
 226                         spin_unlock_irqrestore(&rq->lock, flags);
 227                         ev.device = qp->ibqp.device;
 228                         ev.element.srq = qp->ibqp.srq;
 229                         ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
 230                         handler(&ev, srq->ibsrq.srq_context);
 231                         goto bail;
 232                 }
 233         }
 234 unlock:
 235         spin_unlock_irqrestore(&rq->lock, flags);
 236 bail:
 237         return ret;
 238 }
 239
 240 /**
 241  * ipath_ruc_loopback - handle UC and RC lookback requests
 242  * @sqp: the sending QP
 243  *
 244  * This is called from ipath_do_send() to
 245  * forward a WQE addressed to the same HCA.
 246  * Note that although we are single threaded due to the tasklet, we still
 247  * have to protect against post_send().  We don't have to worry about
 248  * receive interrupts since this is a connected protocol and all packets
 249  * will pass through here.
 250  */
 251 static void ipath_ruc_loopback(struct ipath_qp *sqp)
 252 {
 253         struct ipath_ibdev *dev = to_idev(sqp->ibqp.device);
 254         struct ipath_qp *qp;
 255         struct ipath_swqe *wqe;
 256         struct ipath_sge *sge;
 257         unsigned long flags;
 258         struct ib_wc wc;
 259         u64 sdata;
 260         atomic64_t *maddr;
 261         enum ib_wc_status send_status;
 262
 263         /*
 264          * Note that we check the responder QP state after
 265          * checking the requester's state.
 266          */
 267         qp = ipath_lookup_qpn(&dev->qp_table, sqp->remote_qpn);
 268
 269         spin_lock_irqsave(&sqp->s_lock, flags);
 270
 271         /* Return if we are already busy processing a work request. */
 272         if ((sqp->s_flags & (IPATH_S_BUSY | IPATH_S_ANY_WAIT)) ||
 273             !(ib_ipath_state_ops[sqp->state] & IPATH_PROCESS_OR_FLUSH_SEND))
 274                 goto unlock;
 275
 276         sqp->s_flags |= IPATH_S_BUSY;
 277
 278 again:
 279         if (sqp->s_last == sqp->s_head)
 280                 goto clr_busy;
 281         wqe = get_swqe_ptr(sqp, sqp->s_last);
 282
 283         /* Return if it is not OK to start a new work reqeust. */
 284         if (!(ib_ipath_state_ops[sqp->state] & IPATH_PROCESS_NEXT_SEND_OK)) {
 285                 if (!(ib_ipath_state_ops[sqp->state] & IPATH_FLUSH_SEND))
 286                         goto clr_busy;
 287                 /* We are in the error state, flush the work request. */
 288                 send_status = IB_WC_WR_FLUSH_ERR;
 289                 goto flush_send;
 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          * We increment s_cur to indicate s_last is in progress.
 296          */
 297         if (sqp->s_last == sqp->s_cur) {
 298                 if (++sqp->s_cur >= sqp->s_size)
 299                         sqp->s_cur = 0;
 300         }
 301         spin_unlock_irqrestore(&sqp->s_lock, flags);
 302
 303         if (!qp || !(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_RECV_OK)) {
 304                 dev->n_pkt_drops++;
 305                 /*
 306                  * For RC, the requester would timeout and retry so
 307                  * shortcut the timeouts and just signal too many retries.
 308                  */
 309                 if (sqp->ibqp.qp_type == IB_QPT_RC)
 310                         send_status = IB_WC_RETRY_EXC_ERR;
 311                 else
 312                         send_status = IB_WC_SUCCESS;
 313                 goto serr;
 314         }
 315
 316         memset(&wc, 0, sizeof wc);
 317         send_status = IB_WC_SUCCESS;
 318
 319         sqp->s_sge.sge = wqe->sg_list[0];
 320         sqp->s_sge.sg_list = wqe->sg_list + 1;
 321         sqp->s_sge.num_sge = wqe->wr.num_sge;
 322         sqp->s_len = wqe->length;
 323         switch (wqe->wr.opcode) {
 324         case IB_WR_SEND_WITH_IMM:
 325                 wc.wc_flags = IB_WC_WITH_IMM;
 326                 wc.ex.imm_data = wqe->wr.ex.imm_data;
 327                 /* FALLTHROUGH */
 328         case IB_WR_SEND:
 329                 if (!ipath_get_rwqe(qp, 0))
 330                         goto rnr_nak;
 331                 break;
 332
 333         case IB_WR_RDMA_WRITE_WITH_IMM:
 334                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
 335                         goto inv_err;
 336                 wc.wc_flags = IB_WC_WITH_IMM;
 337                 wc.ex.imm_data = wqe->wr.ex.imm_data;
 338                 if (!ipath_get_rwqe(qp, 1))
 339                         goto rnr_nak;
 340                 /* FALLTHROUGH */
 341         case IB_WR_RDMA_WRITE:
 342                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
 343                         goto inv_err;
 344                 if (wqe->length == 0)
 345                         break;
 346                 if (unlikely(!ipath_rkey_ok(qp, &qp->r_sge, wqe->length,
 347                                             wqe->wr.wr.rdma.remote_addr,
 348                                             wqe->wr.wr.rdma.rkey,
 349                                             IB_ACCESS_REMOTE_WRITE)))
 350                         goto acc_err;
 351                 break;
 352
 353         case IB_WR_RDMA_READ:
 354                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ)))
 355                         goto inv_err;
 356                 if (unlikely(!ipath_rkey_ok(qp, &sqp->s_sge, wqe->length,
 357                                             wqe->wr.wr.rdma.remote_addr,
 358                                             wqe->wr.wr.rdma.rkey,
 359                                             IB_ACCESS_REMOTE_READ)))
 360                         goto acc_err;
 361                 qp->r_sge.sge = wqe->sg_list[0];
 362                 qp->r_sge.sg_list = wqe->sg_list + 1;
 363                 qp->r_sge.num_sge = wqe->wr.num_sge;
 364                 break;
 365
 366         case IB_WR_ATOMIC_CMP_AND_SWP:
 367         case IB_WR_ATOMIC_FETCH_AND_ADD:
 368                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC)))
 369                         goto inv_err;
 370                 if (unlikely(!ipath_rkey_ok(qp, &qp->r_sge, sizeof(u64),
 371                                             wqe->wr.wr.atomic.remote_addr,
 372                                             wqe->wr.wr.atomic.rkey,
 373                                             IB_ACCESS_REMOTE_ATOMIC)))
 374                         goto acc_err;
 375                 /* Perform atomic OP and save result. */
 376                 maddr = (atomic64_t *) qp->r_sge.sge.vaddr;
 377                 sdata = wqe->wr.wr.atomic.compare_add;
 378                 *(u64 *) sqp->s_sge.sge.vaddr =
 379                         (wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) ?
 380                         (u64) atomic64_add_return(sdata, maddr) - sdata :
 381                         (u64) cmpxchg((u64 *) qp->r_sge.sge.vaddr,
 382                                       sdata, wqe->wr.wr.atomic.swap);
 383                 goto send_comp;
 384
 385         default:
 386                 send_status = IB_WC_LOC_QP_OP_ERR;
 387                 goto serr;
 388         }
 389
 390         sge = &sqp->s_sge.sge;
 391         while (sqp->s_len) {
 392                 u32 len = sqp->s_len;
 393
 394                 if (len > sge->length)
 395                         len = sge->length;
 396                 if (len > sge->sge_length)
 397                         len = sge->sge_length;
 398                 BUG_ON(len == 0);
 399                 ipath_copy_sge(&qp->r_sge, sge->vaddr, len);
 400                 sge->vaddr += len;
 401                 sge->length -= len;
 402                 sge->sge_length -= len;
 403                 if (sge->sge_length == 0) {
 404                         if (--sqp->s_sge.num_sge)
 405                                 *sge = *sqp->s_sge.sg_list++;
 406                 } else if (sge->length == 0 && sge->mr != NULL) {
 407                         if (++sge->n >= IPATH_SEGSZ) {
 408                                 if (++sge->m >= sge->mr->mapsz)
 409                                         break;
 410                                 sge->n = 0;
 411                         }
 412                         sge->vaddr =
 413                                 sge->mr->map[sge->m]->segs[sge->n].vaddr;
 414                         sge->length =
 415                                 sge->mr->map[sge->m]->segs[sge->n].length;
 416                 }
 417                 sqp->s_len -= len;
 418         }
 419
 420         if (!test_and_clear_bit(IPATH_R_WRID_VALID, &qp->r_aflags))
 421                 goto send_comp;
 422
 423         if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM)
 424                 wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
 425         else
 426                 wc.opcode = IB_WC_RECV;
 427         wc.wr_id = qp->r_wr_id;
 428         wc.status = IB_WC_SUCCESS;
 429         wc.byte_len = wqe->length;
 430         wc.qp = &qp->ibqp;
 431         wc.src_qp = qp->remote_qpn;
 432         wc.slid = qp->remote_ah_attr.dlid;
 433         wc.sl = qp->remote_ah_attr.sl;
 434         wc.port_num = 1;
 435         /* Signal completion event if the solicited bit is set. */
 436         ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
 437                        wqe->wr.send_flags & IB_SEND_SOLICITED);
 438
 439 send_comp:
 440         spin_lock_irqsave(&sqp->s_lock, flags);
 441 flush_send:
 442         sqp->s_rnr_retry = sqp->s_rnr_retry_cnt;
 443         ipath_send_complete(sqp, wqe, send_status);
 444         goto again;
 445
 446 rnr_nak:
 447         /* Handle RNR NAK */
 448         if (qp->ibqp.qp_type == IB_QPT_UC)
 449                 goto send_comp;
 450         /*
 451          * Note: we don't need the s_lock held since the BUSY flag
 452          * makes this single threaded.
 453          */
 454         if (sqp->s_rnr_retry == 0) {
 455                 send_status = IB_WC_RNR_RETRY_EXC_ERR;
 456                 goto serr;
 457         }
 458         if (sqp->s_rnr_retry_cnt < 7)
 459                 sqp->s_rnr_retry--;
 460         spin_lock_irqsave(&sqp->s_lock, flags);
 461         if (!(ib_ipath_state_ops[sqp->state] & IPATH_PROCESS_RECV_OK))
 462                 goto clr_busy;
 463         sqp->s_flags |= IPATH_S_WAITING;
 464         dev->n_rnr_naks++;
 465         sqp->s_rnr_timeout = ib_ipath_rnr_table[qp->r_min_rnr_timer];
 466         ipath_insert_rnr_queue(sqp);
 467         goto clr_busy;
 468
 469 inv_err:
 470         send_status = IB_WC_REM_INV_REQ_ERR;
 471         wc.status = IB_WC_LOC_QP_OP_ERR;
 472         goto err;
 473
 474 acc_err:
 475         send_status = IB_WC_REM_ACCESS_ERR;
 476         wc.status = IB_WC_LOC_PROT_ERR;
 477 err:
 478         /* responder goes to error state */
 479         ipath_rc_error(qp, wc.status);
 480
 481 serr:
 482         spin_lock_irqsave(&sqp->s_lock, flags);
 483         ipath_send_complete(sqp, wqe, send_status);
 484         if (sqp->ibqp.qp_type == IB_QPT_RC) {
 485                 int lastwqe = ipath_error_qp(sqp, IB_WC_WR_FLUSH_ERR);
 486
 487                 sqp->s_flags &= ~IPATH_S_BUSY;
 488                 spin_unlock_irqrestore(&sqp->s_lock, flags);
 489                 if (lastwqe) {
 490                         struct ib_event ev;
 491
 492                         ev.device = sqp->ibqp.device;
 493                         ev.element.qp = &sqp->ibqp;
 494                         ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
 495                         sqp->ibqp.event_handler(&ev, sqp->ibqp.qp_context);
 496                 }
 497                 goto done;
 498         }
 499 clr_busy:
 500         sqp->s_flags &= ~IPATH_S_BUSY;
 501 unlock:
 502         spin_unlock_irqrestore(&sqp->s_lock, flags);
 503 done:
 504         if (qp && atomic_dec_and_test(&qp->refcount))
 505                 wake_up(&qp->wait);
 506 }
 507
 508 static void want_buffer(struct ipath_devdata *dd, struct ipath_qp *qp)
 509 {
 510         if (!(dd->ipath_flags & IPATH_HAS_SEND_DMA) ||
 511             qp->ibqp.qp_type == IB_QPT_SMI) {
 512                 unsigned long flags;
 513
 514                 spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
 515                 dd->ipath_sendctrl |= INFINIPATH_S_PIOINTBUFAVAIL;
 516                 ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
 517                                  dd->ipath_sendctrl);
 518                 ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
 519                 spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
 520         }
 521 }
 522
 523 /**
 524  * ipath_no_bufs_available - tell the layer driver we need buffers
 525  * @qp: the QP that caused the problem
 526  * @dev: the device we ran out of buffers on
 527  *
 528  * Called when we run out of PIO buffers.
 529  * If we are now in the error state, return zero to flush the
 530  * send work request.
 531  */
 532 static int ipath_no_bufs_available(struct ipath_qp *qp,
 533                                     struct ipath_ibdev *dev)
 534 {
 535         unsigned long flags;
 536         int ret = 1;
 537
 538         /*
 539          * Note that as soon as want_buffer() is called and
 540          * possibly before it returns, ipath_ib_piobufavail()
 541          * could be called. Therefore, put QP on the piowait list before
 542          * enabling the PIO avail interrupt.
 543          */
 544         spin_lock_irqsave(&qp->s_lock, flags);
 545         if (ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK) {
 546                 dev->n_piowait++;
 547                 qp->s_flags |= IPATH_S_WAITING;
 548                 qp->s_flags &= ~IPATH_S_BUSY;
 549                 spin_lock(&dev->pending_lock);
 550                 if (list_empty(&qp->piowait))
 551                         list_add_tail(&qp->piowait, &dev->piowait);
 552                 spin_unlock(&dev->pending_lock);
 553         } else
 554                 ret = 0;
 555         spin_unlock_irqrestore(&qp->s_lock, flags);
 556         if (ret)
 557                 want_buffer(dev->dd, qp);
 558         return ret;
 559 }
 560
 561 /**
 562  * ipath_make_grh - construct a GRH header
 563  * @dev: a pointer to the ipath device
 564  * @hdr: a pointer to the GRH header being constructed
 565  * @grh: the global route address to send to
 566  * @hwords: the number of 32 bit words of header being sent
 567  * @nwords: the number of 32 bit words of data being sent
 568  *
 569  * Return the size of the header in 32 bit words.
 570  */
 571 u32 ipath_make_grh(struct ipath_ibdev *dev, struct ib_grh *hdr,
 572                    struct ib_global_route *grh, u32 hwords, u32 nwords)
 573 {
 574         hdr->version_tclass_flow =
 575                 cpu_to_be32((6 << 28) |
 576                             (grh->traffic_class << 20) |
 577                             grh->flow_label);
 578         hdr->paylen = cpu_to_be16((hwords - 2 + nwords + SIZE_OF_CRC) << 2);
 579         /* next_hdr is defined by C8-7 in ch. 8.4.1 */
 580         hdr->next_hdr = 0x1B;
 581         hdr->hop_limit = grh->hop_limit;
 582         /* The SGID is 32-bit aligned. */
 583         hdr->sgid.global.subnet_prefix = dev->gid_prefix;
 584         hdr->sgid.global.interface_id = dev->dd->ipath_guid;
 585         hdr->dgid = grh->dgid;
 586
 587         /* GRH header size in 32-bit words. */
 588         return sizeof(struct ib_grh) / sizeof(u32);
 589 }
 590
 591 void ipath_make_ruc_header(struct ipath_ibdev *dev, struct ipath_qp *qp,
 592                            struct ipath_other_headers *ohdr,
 593                            u32 bth0, u32 bth2)
 594 {
 595         u16 lrh0;
 596         u32 nwords;
 597         u32 extra_bytes;
 598
 599         /* Construct the header. */
 600         extra_bytes = -qp->s_cur_size & 3;
 601         nwords = (qp->s_cur_size + extra_bytes) >> 2;
 602         lrh0 = IPATH_LRH_BTH;
 603         if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
 604                 qp->s_hdrwords += ipath_make_grh(dev, &qp->s_hdr.u.l.grh,
 605                                                  &qp->remote_ah_attr.grh,
 606                                                  qp->s_hdrwords, nwords);
 607                 lrh0 = IPATH_LRH_GRH;
 608         }
 609         lrh0 |= qp->remote_ah_attr.sl << 4;
 610         qp->s_hdr.lrh[0] = cpu_to_be16(lrh0);
 611         qp->s_hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
 612         qp->s_hdr.lrh[2] = cpu_to_be16(qp->s_hdrwords + nwords + SIZE_OF_CRC);
 613         qp->s_hdr.lrh[3] = cpu_to_be16(dev->dd->ipath_lid |
 614                                        qp->remote_ah_attr.src_path_bits);
 615         bth0 |= ipath_get_pkey(dev->dd, qp->s_pkey_index);
 616         bth0 |= extra_bytes << 20;
 617         ohdr->bth[0] = cpu_to_be32(bth0 | (1 << 22));
 618         ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
 619         ohdr->bth[2] = cpu_to_be32(bth2);
 620 }
 621
 622 /**
 623  * ipath_do_send - perform a send on a QP
 624  * @data: contains a pointer to the QP
 625  *
 626  * Process entries in the send work queue until credit or queue is
 627  * exhausted.  Only allow one CPU to send a packet per QP (tasklet).
 628  * Otherwise, two threads could send packets out of order.
 629  */
 630 void ipath_do_send(unsigned long data)
 631 {
 632         struct ipath_qp *qp = (struct ipath_qp *)data;
 633         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
 634         int (*make_req)(struct ipath_qp *qp);
 635         unsigned long flags;
 636
 637         if ((qp->ibqp.qp_type == IB_QPT_RC ||
 638              qp->ibqp.qp_type == IB_QPT_UC) &&
 639             qp->remote_ah_attr.dlid == dev->dd->ipath_lid) {
 640                 ipath_ruc_loopback(qp);
 641                 goto bail;
 642         }
 643
 644         if (qp->ibqp.qp_type == IB_QPT_RC)
 645                make_req = ipath_make_rc_req;
 646         else if (qp->ibqp.qp_type == IB_QPT_UC)
 647                make_req = ipath_make_uc_req;
 648         else
 649                make_req = ipath_make_ud_req;
 650
 651         spin_lock_irqsave(&qp->s_lock, flags);
 652
 653         /* Return if we are already busy processing a work request. */
 654         if ((qp->s_flags & (IPATH_S_BUSY | IPATH_S_ANY_WAIT)) ||
 655             !(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_OR_FLUSH_SEND)) {
 656                 spin_unlock_irqrestore(&qp->s_lock, flags);
 657                 goto bail;
 658         }
 659
 660         qp->s_flags |= IPATH_S_BUSY;
 661
 662         spin_unlock_irqrestore(&qp->s_lock, flags);
 663
 664 again:
 665         /* Check for a constructed packet to be sent. */
 666         if (qp->s_hdrwords != 0) {
 667                 /*
 668                  * If no PIO bufs are available, return.  An interrupt will
 669                  * call ipath_ib_piobufavail() when one is available.
 670                  */
 671                 if (ipath_verbs_send(qp, &qp->s_hdr, qp->s_hdrwords,
 672                                      qp->s_cur_sge, qp->s_cur_size)) {
 673                         if (ipath_no_bufs_available(qp, dev))
 674                                 goto bail;
 675                 }
 676                 dev->n_unicast_xmit++;
 677                 /* Record that we sent the packet and s_hdr is empty. */
 678                 qp->s_hdrwords = 0;
 679         }
 680
 681         if (make_req(qp))
 682                 goto again;
 683
 684 bail:;
 685 }
 686
 687 /*
 688  * This should be called with s_lock held.
 689  */
 690 void ipath_send_complete(struct ipath_qp *qp, struct ipath_swqe *wqe,
 691                          enum ib_wc_status status)
 692 {
 693         u32 old_last, last;
 694
 695         if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_OR_FLUSH_SEND))
 696                 return;
 697
 698         /* See ch. 11.2.4.1 and 10.7.3.1 */
 699         if (!(qp->s_flags & IPATH_S_SIGNAL_REQ_WR) ||
 700             (wqe->wr.send_flags & IB_SEND_SIGNALED) ||
 701             status != IB_WC_SUCCESS) {
 702                 struct ib_wc wc;
 703
 704                 memset(&wc, 0, sizeof wc);
 705                 wc.wr_id = wqe->wr.wr_id;
 706                 wc.status = status;
 707                 wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
 708                 wc.qp = &qp->ibqp;
 709                 if (status == IB_WC_SUCCESS)
 710                         wc.byte_len = wqe->length;
 711                 ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc,
 712                                status != IB_WC_SUCCESS);
 713         }
 714
 715         old_last = last = qp->s_last;
 716         if (++last >= qp->s_size)
 717                 last = 0;
 718         qp->s_last = last;
 719         if (qp->s_cur == old_last)
 720                 qp->s_cur = last;
 721         if (qp->s_tail == old_last)
 722                 qp->s_tail = last;
 723         if (qp->state == IB_QPS_SQD && last == qp->s_cur)
 724                 qp->s_draining = 0;
 725 }