SAFE public projects git trees. - safe/jmp/linux-2.6/blob - drivers/infiniband/hw/ipath/ipath_ud.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 <rdma/ib_smi.h>
  35
  36 #include "ipath_verbs.h"
  37 #include "ipath_kernel.h"
  38
  39 static int init_sge(struct ipath_qp *qp, struct ipath_rwqe *wqe,
  40                     u32 *lengthp, struct ipath_sge_state *ss)
  41 {
  42         int user = to_ipd(qp->ibqp.pd)->user;
  43         int i, j, ret;
  44         struct ib_wc wc;
  45
  46         *lengthp = 0;
  47         for (i = j = 0; i < wqe->num_sge; i++) {
  48                 if (wqe->sg_list[i].length == 0)
  49                         continue;
  50                 /* Check LKEY */
  51                 if ((user && wqe->sg_list[i].lkey == 0) ||
  52                     !ipath_lkey_ok(qp, j ? &ss->sg_list[j - 1] : &ss->sge,
  53                                    &wqe->sg_list[i], IB_ACCESS_LOCAL_WRITE))
  54                         goto bad_lkey;
  55                 *lengthp += wqe->sg_list[i].length;
  56                 j++;
  57         }
  58         ss->num_sge = j;
  59         ret = 1;
  60         goto bail;
  61
  62 bad_lkey:
  63         wc.wr_id = wqe->wr_id;
  64         wc.status = IB_WC_LOC_PROT_ERR;
  65         wc.opcode = IB_WC_RECV;
  66         wc.vendor_err = 0;
  67         wc.byte_len = 0;
  68         wc.imm_data = 0;
  69         wc.qp = &qp->ibqp;
  70         wc.src_qp = 0;
  71         wc.wc_flags = 0;
  72         wc.pkey_index = 0;
  73         wc.slid = 0;
  74         wc.sl = 0;
  75         wc.dlid_path_bits = 0;
  76         wc.port_num = 0;
  77         /* Signal solicited completion event. */
  78         ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
  79         ret = 0;
  80 bail:
  81         return ret;
  82 }
  83
  84 /**
  85  * ipath_ud_loopback - handle send on loopback QPs
  86  * @sqp: the QP
  87  * @ss: the SGE state
  88  * @length: the length of the data to send
  89  * @wr: the work request
  90  * @wc: the work completion entry
  91  *
  92  * This is called from ipath_post_ud_send() to forward a WQE addressed
  93  * to the same HCA.
  94  * Note that the receive interrupt handler may be calling ipath_ud_rcv()
  95  * while this is being called.
  96  */
  97 static void ipath_ud_loopback(struct ipath_qp *sqp,
  98                               struct ipath_sge_state *ss,
  99                               u32 length, struct ib_send_wr *wr,
 100                               struct ib_wc *wc)
 101 {
 102         struct ipath_ibdev *dev = to_idev(sqp->ibqp.device);
 103         struct ipath_qp *qp;
 104         struct ib_ah_attr *ah_attr;
 105         unsigned long flags;
 106         struct ipath_rq *rq;
 107         struct ipath_srq *srq;
 108         struct ipath_sge_state rsge;
 109         struct ipath_sge *sge;
 110         struct ipath_rwq *wq;
 111         struct ipath_rwqe *wqe;
 112         void (*handler)(struct ib_event *, void *);
 113         u32 tail;
 114         u32 rlen;
 115
 116         qp = ipath_lookup_qpn(&dev->qp_table, wr->wr.ud.remote_qpn);
 117         if (!qp)
 118                 return;
 119
 120         /*
 121          * Check that the qkey matches (except for QP0, see 9.6.1.4.1).
 122          * Qkeys with the high order bit set mean use the
 123          * qkey from the QP context instead of the WR (see 10.2.5).
 124          */
 125         if (unlikely(qp->ibqp.qp_num &&
 126                      ((int) wr->wr.ud.remote_qkey < 0
 127                       ? qp->qkey : wr->wr.ud.remote_qkey) != qp->qkey)) {
 128                 /* XXX OK to lose a count once in a while. */
 129                 dev->qkey_violations++;
 130                 dev->n_pkt_drops++;
 131                 goto done;
 132         }
 133
 134         /*
 135          * A GRH is expected to preceed the data even if not
 136          * present on the wire.
 137          */
 138         wc->byte_len = length + sizeof(struct ib_grh);
 139
 140         if (wr->opcode == IB_WR_SEND_WITH_IMM) {
 141                 wc->wc_flags = IB_WC_WITH_IMM;
 142                 wc->imm_data = wr->imm_data;
 143         } else {
 144                 wc->wc_flags = 0;
 145                 wc->imm_data = 0;
 146         }
 147
 148         if (wr->num_sge > 1) {
 149                 rsge.sg_list = kmalloc((wr->num_sge - 1) *
 150                                         sizeof(struct ipath_sge),
 151                                        GFP_ATOMIC);
 152         } else
 153                 rsge.sg_list = NULL;
 154
 155         /*
 156          * Get the next work request entry to find where to put the data.
 157          * Note that it is safe to drop the lock after changing rq->tail
 158          * since ipath_post_receive() won't fill the empty slot.
 159          */
 160         if (qp->ibqp.srq) {
 161                 srq = to_isrq(qp->ibqp.srq);
 162                 handler = srq->ibsrq.event_handler;
 163                 rq = &srq->rq;
 164         } else {
 165                 srq = NULL;
 166                 handler = NULL;
 167                 rq = &qp->r_rq;
 168         }
 169
 170         spin_lock_irqsave(&rq->lock, flags);
 171         wq = rq->wq;
 172         tail = wq->tail;
 173         while (1) {
 174                 if (unlikely(tail == wq->head)) {
 175                         spin_unlock_irqrestore(&rq->lock, flags);
 176                         dev->n_pkt_drops++;
 177                         goto bail_sge;
 178                 }
 179                 wqe = get_rwqe_ptr(rq, tail);
 180                 if (++tail >= rq->size)
 181                         tail = 0;
 182                 if (init_sge(qp, wqe, &rlen, &rsge))
 183                         break;
 184                 wq->tail = tail;
 185         }
 186         /* Silently drop packets which are too big. */
 187         if (wc->byte_len > rlen) {
 188                 spin_unlock_irqrestore(&rq->lock, flags);
 189                 dev->n_pkt_drops++;
 190                 goto bail_sge;
 191         }
 192         wq->tail = tail;
 193         wc->wr_id = wqe->wr_id;
 194         if (handler) {
 195                 u32 n;
 196
 197                 /*
 198                  * validate head pointer value and compute
 199                  * the number of remaining WQEs.
 200                  */
 201                 n = wq->head;
 202                 if (n >= rq->size)
 203                         n = 0;
 204                 if (n < tail)
 205                         n += rq->size - tail;
 206                 else
 207                         n -= tail;
 208                 if (n < srq->limit) {
 209                         struct ib_event ev;
 210
 211                         srq->limit = 0;
 212                         spin_unlock_irqrestore(&rq->lock, flags);
 213                         ev.device = qp->ibqp.device;
 214                         ev.element.srq = qp->ibqp.srq;
 215                         ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
 216                         handler(&ev, srq->ibsrq.srq_context);
 217                 } else
 218                         spin_unlock_irqrestore(&rq->lock, flags);
 219         } else
 220                 spin_unlock_irqrestore(&rq->lock, flags);
 221
 222         ah_attr = &to_iah(wr->wr.ud.ah)->attr;
 223         if (ah_attr->ah_flags & IB_AH_GRH) {
 224                 ipath_copy_sge(&rsge, &ah_attr->grh, sizeof(struct ib_grh));
 225                 wc->wc_flags |= IB_WC_GRH;
 226         } else
 227                 ipath_skip_sge(&rsge, sizeof(struct ib_grh));
 228         sge = &ss->sge;
 229         while (length) {
 230                 u32 len = sge->length;
 231
 232                 if (len > length)
 233                         len = length;
 234                 if (len > sge->sge_length)
 235                         len = sge->sge_length;
 236                 BUG_ON(len == 0);
 237                 ipath_copy_sge(&rsge, sge->vaddr, len);
 238                 sge->vaddr += len;
 239                 sge->length -= len;
 240                 sge->sge_length -= len;
 241                 if (sge->sge_length == 0) {
 242                         if (--ss->num_sge)
 243                                 *sge = *ss->sg_list++;
 244                 } else if (sge->length == 0 && sge->mr != NULL) {
 245                         if (++sge->n >= IPATH_SEGSZ) {
 246                                 if (++sge->m >= sge->mr->mapsz)
 247                                         break;
 248                                 sge->n = 0;
 249                         }
 250                         sge->vaddr =
 251                                 sge->mr->map[sge->m]->segs[sge->n].vaddr;
 252                         sge->length =
 253                                 sge->mr->map[sge->m]->segs[sge->n].length;
 254                 }
 255                 length -= len;
 256         }
 257         wc->status = IB_WC_SUCCESS;
 258         wc->opcode = IB_WC_RECV;
 259         wc->vendor_err = 0;
 260         wc->qp = &qp->ibqp;
 261         wc->src_qp = sqp->ibqp.qp_num;
 262         /* XXX do we know which pkey matched? Only needed for GSI. */
 263         wc->pkey_index = 0;
 264         wc->slid = dev->dd->ipath_lid |
 265                 (ah_attr->src_path_bits &
 266                  ((1 << (dev->mkeyprot_resv_lmc & 7)) - 1));
 267         wc->sl = ah_attr->sl;
 268         wc->dlid_path_bits =
 269                 ah_attr->dlid & ((1 << (dev->mkeyprot_resv_lmc & 7)) - 1);
 270         /* Signal completion event if the solicited bit is set. */
 271         ipath_cq_enter(to_icq(qp->ibqp.recv_cq), wc,
 272                        wr->send_flags & IB_SEND_SOLICITED);
 273
 274 bail_sge:
 275         kfree(rsge.sg_list);
 276 done:
 277         if (atomic_dec_and_test(&qp->refcount))
 278                 wake_up(&qp->wait);
 279 }
 280
 281 /**
 282  * ipath_post_ud_send - post a UD send on QP
 283  * @qp: the QP
 284  * @wr: the work request
 285  *
 286  * Note that we actually send the data as it is posted instead of putting
 287  * the request into a ring buffer.  If we wanted to use a ring buffer,
 288  * we would need to save a reference to the destination address in the SWQE.
 289  */
 290 int ipath_post_ud_send(struct ipath_qp *qp, struct ib_send_wr *wr)
 291 {
 292         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
 293         struct ipath_other_headers *ohdr;
 294         struct ib_ah_attr *ah_attr;
 295         struct ipath_sge_state ss;
 296         struct ipath_sge *sg_list;
 297         struct ib_wc wc;
 298         u32 hwords;
 299         u32 nwords;
 300         u32 len;
 301         u32 extra_bytes;
 302         u32 bth0;
 303         u16 lrh0;
 304         u16 lid;
 305         int i;
 306         int ret;
 307
 308         if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK)) {
 309                 ret = 0;
 310                 goto bail;
 311         }
 312
 313         if (wr->wr.ud.ah->pd != qp->ibqp.pd) {
 314                 ret = -EPERM;
 315                 goto bail;
 316         }
 317
 318         /* IB spec says that num_sge == 0 is OK. */
 319         if (wr->num_sge > qp->s_max_sge) {
 320                 ret = -EINVAL;
 321                 goto bail;
 322         }
 323
 324         if (wr->num_sge > 1) {
 325                 sg_list = kmalloc((qp->s_max_sge - 1) * sizeof(*sg_list),
 326                                   GFP_ATOMIC);
 327                 if (!sg_list) {
 328                         ret = -ENOMEM;
 329                         goto bail;
 330                 }
 331         } else
 332                 sg_list = NULL;
 333
 334         /* Check the buffer to send. */
 335         ss.sg_list = sg_list;
 336         ss.sge.mr = NULL;
 337         ss.sge.vaddr = NULL;
 338         ss.sge.length = 0;
 339         ss.sge.sge_length = 0;
 340         ss.num_sge = 0;
 341         len = 0;
 342         for (i = 0; i < wr->num_sge; i++) {
 343                 /* Check LKEY */
 344                 if (to_ipd(qp->ibqp.pd)->user && wr->sg_list[i].lkey == 0) {
 345                         ret = -EINVAL;
 346                         goto bail;
 347                 }
 348
 349                 if (wr->sg_list[i].length == 0)
 350                         continue;
 351                 if (!ipath_lkey_ok(qp, ss.num_sge ?
 352                                    sg_list + ss.num_sge - 1 : &ss.sge,
 353                                    &wr->sg_list[i], 0)) {
 354                         ret = -EINVAL;
 355                         goto bail;
 356                 }
 357                 len += wr->sg_list[i].length;
 358                 ss.num_sge++;
 359         }
 360         /* Check for invalid packet size. */
 361         if (len > dev->dd->ipath_ibmtu) {
 362                 ret = -EINVAL;
 363                 goto bail;
 364         }
 365         extra_bytes = (4 - len) & 3;
 366         nwords = (len + extra_bytes) >> 2;
 367
 368         /* Construct the header. */
 369         ah_attr = &to_iah(wr->wr.ud.ah)->attr;
 370         if (ah_attr->dlid == 0) {
 371                 ret = -EINVAL;
 372                 goto bail;
 373         }
 374         if (ah_attr->dlid >= IPATH_MULTICAST_LID_BASE) {
 375                 if (ah_attr->dlid != IPATH_PERMISSIVE_LID)
 376                         dev->n_multicast_xmit++;
 377                 else
 378                         dev->n_unicast_xmit++;
 379         } else {
 380                 dev->n_unicast_xmit++;
 381                 lid = ah_attr->dlid &
 382                         ~((1 << (dev->mkeyprot_resv_lmc & 7)) - 1);
 383                 if (unlikely(lid == dev->dd->ipath_lid)) {
 384                         /*
 385                          * Pass in an uninitialized ib_wc to save stack
 386                          * space.
 387                          */
 388                         ipath_ud_loopback(qp, &ss, len, wr, &wc);
 389                         goto done;
 390                 }
 391         }
 392         if (ah_attr->ah_flags & IB_AH_GRH) {
 393                 /* Header size in 32-bit words. */
 394                 hwords = 17;
 395                 lrh0 = IPATH_LRH_GRH;
 396                 ohdr = &qp->s_hdr.u.l.oth;
 397                 qp->s_hdr.u.l.grh.version_tclass_flow =
 398                         cpu_to_be32((6 << 28) |
 399                                     (ah_attr->grh.traffic_class << 20) |
 400                                     ah_attr->grh.flow_label);
 401                 qp->s_hdr.u.l.grh.paylen =
 402                         cpu_to_be16(((wr->opcode ==
 403                                       IB_WR_SEND_WITH_IMM ? 6 : 5) +
 404                                      nwords + SIZE_OF_CRC) << 2);
 405                 /* next_hdr is defined by C8-7 in ch. 8.4.1 */
 406                 qp->s_hdr.u.l.grh.next_hdr = 0x1B;
 407                 qp->s_hdr.u.l.grh.hop_limit = ah_attr->grh.hop_limit;
 408                 /* The SGID is 32-bit aligned. */
 409                 qp->s_hdr.u.l.grh.sgid.global.subnet_prefix =
 410                         dev->gid_prefix;
 411                 qp->s_hdr.u.l.grh.sgid.global.interface_id =
 412                         dev->dd->ipath_guid;
 413                 qp->s_hdr.u.l.grh.dgid = ah_attr->grh.dgid;
 414                 /*
 415                  * Don't worry about sending to locally attached multicast
 416                  * QPs.  It is unspecified by the spec. what happens.
 417                  */
 418         } else {
 419                 /* Header size in 32-bit words. */
 420                 hwords = 7;
 421                 lrh0 = IPATH_LRH_BTH;
 422                 ohdr = &qp->s_hdr.u.oth;
 423         }
 424         if (wr->opcode == IB_WR_SEND_WITH_IMM) {
 425                 ohdr->u.ud.imm_data = wr->imm_data;
 426                 wc.imm_data = wr->imm_data;
 427                 hwords += 1;
 428                 bth0 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE << 24;
 429         } else if (wr->opcode == IB_WR_SEND) {
 430                 wc.imm_data = 0;
 431                 bth0 = IB_OPCODE_UD_SEND_ONLY << 24;
 432         } else {
 433                 ret = -EINVAL;
 434                 goto bail;
 435         }
 436         lrh0 |= ah_attr->sl << 4;
 437         if (qp->ibqp.qp_type == IB_QPT_SMI)
 438                 lrh0 |= 0xF000; /* Set VL (see ch. 13.5.3.1) */
 439         qp->s_hdr.lrh[0] = cpu_to_be16(lrh0);
 440         qp->s_hdr.lrh[1] = cpu_to_be16(ah_attr->dlid);  /* DEST LID */
 441         qp->s_hdr.lrh[2] = cpu_to_be16(hwords + nwords + SIZE_OF_CRC);
 442         lid = dev->dd->ipath_lid;
 443         if (lid) {
 444                 lid |= ah_attr->src_path_bits &
 445                         ((1 << (dev->mkeyprot_resv_lmc & 7)) - 1);
 446                 qp->s_hdr.lrh[3] = cpu_to_be16(lid);
 447         } else
 448                 qp->s_hdr.lrh[3] = IB_LID_PERMISSIVE;
 449         if (wr->send_flags & IB_SEND_SOLICITED)
 450                 bth0 |= 1 << 23;
 451         bth0 |= extra_bytes << 20;
 452         bth0 |= qp->ibqp.qp_type == IB_QPT_SMI ? IPATH_DEFAULT_P_KEY :
 453                 ipath_get_pkey(dev->dd, qp->s_pkey_index);
 454         ohdr->bth[0] = cpu_to_be32(bth0);
 455         /*
 456          * Use the multicast QP if the destination LID is a multicast LID.
 457          */
 458         ohdr->bth[1] = ah_attr->dlid >= IPATH_MULTICAST_LID_BASE &&
 459                 ah_attr->dlid != IPATH_PERMISSIVE_LID ?
 460                 __constant_cpu_to_be32(IPATH_MULTICAST_QPN) :
 461                 cpu_to_be32(wr->wr.ud.remote_qpn);
 462         /* XXX Could lose a PSN count but not worth locking */
 463         ohdr->bth[2] = cpu_to_be32(qp->s_next_psn++ & IPATH_PSN_MASK);
 464         /*
 465          * Qkeys with the high order bit set mean use the
 466          * qkey from the QP context instead of the WR (see 10.2.5).
 467          */
 468         ohdr->u.ud.deth[0] = cpu_to_be32((int)wr->wr.ud.remote_qkey < 0 ?
 469                                          qp->qkey : wr->wr.ud.remote_qkey);
 470         ohdr->u.ud.deth[1] = cpu_to_be32(qp->ibqp.qp_num);
 471         if (ipath_verbs_send(dev->dd, hwords, (u32 *) &qp->s_hdr,
 472                              len, &ss))
 473                 dev->n_no_piobuf++;
 474
 475 done:
 476         /* Queue the completion status entry. */
 477         if (!(qp->s_flags & IPATH_S_SIGNAL_REQ_WR) ||
 478             (wr->send_flags & IB_SEND_SIGNALED)) {
 479                 wc.wr_id = wr->wr_id;
 480                 wc.status = IB_WC_SUCCESS;
 481                 wc.vendor_err = 0;
 482                 wc.opcode = IB_WC_SEND;
 483                 wc.byte_len = len;
 484                 wc.qp = &qp->ibqp;
 485                 wc.src_qp = 0;
 486                 wc.wc_flags = 0;
 487                 /* XXX initialize other fields? */
 488                 ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 0);
 489         }
 490         kfree(sg_list);
 491
 492         ret = 0;
 493
 494 bail:
 495         return ret;
 496 }
 497
 498 /**
 499  * ipath_ud_rcv - receive an incoming UD packet
 500  * @dev: the device the packet came in on
 501  * @hdr: the packet header
 502  * @has_grh: true if the packet has a GRH
 503  * @data: the packet data
 504  * @tlen: the packet length
 505  * @qp: the QP the packet came on
 506  *
 507  * This is called from ipath_qp_rcv() to process an incoming UD packet
 508  * for the given QP.
 509  * Called at interrupt level.
 510  */
 511 void ipath_ud_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
 512                   int has_grh, void *data, u32 tlen, struct ipath_qp *qp)
 513 {
 514         struct ipath_other_headers *ohdr;
 515         int opcode;
 516         u32 hdrsize;
 517         u32 pad;
 518         struct ib_wc wc;
 519         u32 qkey;
 520         u32 src_qp;
 521         u16 dlid;
 522         int header_in_data;
 523
 524         /* Check for GRH */
 525         if (!has_grh) {
 526                 ohdr = &hdr->u.oth;
 527                 hdrsize = 8 + 12 + 8;   /* LRH + BTH + DETH */
 528                 qkey = be32_to_cpu(ohdr->u.ud.deth[0]);
 529                 src_qp = be32_to_cpu(ohdr->u.ud.deth[1]);
 530                 header_in_data = 0;
 531         } else {
 532                 ohdr = &hdr->u.l.oth;
 533                 hdrsize = 8 + 40 + 12 + 8; /* LRH + GRH + BTH + DETH */
 534                 /*
 535                  * The header with GRH is 68 bytes and the core driver sets
 536                  * the eager header buffer size to 56 bytes so the last 12
 537                  * bytes of the IB header is in the data buffer.
 538                  */
 539                 header_in_data = dev->dd->ipath_rcvhdrentsize == 16;
 540                 if (header_in_data) {
 541                         qkey = be32_to_cpu(((__be32 *) data)[1]);
 542                         src_qp = be32_to_cpu(((__be32 *) data)[2]);
 543                         data += 12;
 544                 } else {
 545                         qkey = be32_to_cpu(ohdr->u.ud.deth[0]);
 546                         src_qp = be32_to_cpu(ohdr->u.ud.deth[1]);
 547                 }
 548         }
 549         src_qp &= IPATH_QPN_MASK;
 550
 551         /*
 552          * Check that the permissive LID is only used on QP0
 553          * and the QKEY matches (see 9.6.1.4.1 and 9.6.1.5.1).
 554          */
 555         if (qp->ibqp.qp_num) {
 556                 if (unlikely(hdr->lrh[1] == IB_LID_PERMISSIVE ||
 557                              hdr->lrh[3] == IB_LID_PERMISSIVE)) {
 558                         dev->n_pkt_drops++;
 559                         goto bail;
 560                 }
 561                 if (unlikely(qkey != qp->qkey)) {
 562                         /* XXX OK to lose a count once in a while. */
 563                         dev->qkey_violations++;
 564                         dev->n_pkt_drops++;
 565                         goto bail;
 566                 }
 567         } else if (hdr->lrh[1] == IB_LID_PERMISSIVE ||
 568                    hdr->lrh[3] == IB_LID_PERMISSIVE) {
 569                 struct ib_smp *smp = (struct ib_smp *) data;
 570
 571                 if (smp->mgmt_class != IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
 572                         dev->n_pkt_drops++;
 573                         goto bail;
 574                 }
 575         }
 576
 577         /* Get the number of bytes the message was padded by. */
 578         pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
 579         if (unlikely(tlen < (hdrsize + pad + 4))) {
 580                 /* Drop incomplete packets. */
 581                 dev->n_pkt_drops++;
 582                 goto bail;
 583         }
 584         tlen -= hdrsize + pad + 4;
 585
 586         /* Drop invalid MAD packets (see 13.5.3.1). */
 587         if (unlikely((qp->ibqp.qp_num == 0 &&
 588                       (tlen != 256 ||
 589                        (be16_to_cpu(hdr->lrh[0]) >> 12) != 15)) ||
 590                      (qp->ibqp.qp_num == 1 &&
 591                       (tlen != 256 ||
 592                        (be16_to_cpu(hdr->lrh[0]) >> 12) == 15)))) {
 593                 dev->n_pkt_drops++;
 594                 goto bail;
 595         }
 596
 597         /*
 598          * A GRH is expected to preceed the data even if not
 599          * present on the wire.
 600          */
 601         wc.byte_len = tlen + sizeof(struct ib_grh);
 602
 603         /*
 604          * The opcode is in the low byte when its in network order
 605          * (top byte when in host order).
 606          */
 607         opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
 608         if (qp->ibqp.qp_num > 1 &&
 609             opcode == IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE) {
 610                 if (header_in_data) {
 611                         wc.imm_data = *(__be32 *) data;
 612                         data += sizeof(__be32);
 613                 } else
 614                         wc.imm_data = ohdr->u.ud.imm_data;
 615                 wc.wc_flags = IB_WC_WITH_IMM;
 616                 hdrsize += sizeof(u32);
 617         } else if (opcode == IB_OPCODE_UD_SEND_ONLY) {
 618                 wc.imm_data = 0;
 619                 wc.wc_flags = 0;
 620         } else {
 621                 dev->n_pkt_drops++;
 622                 goto bail;
 623         }
 624
 625         /*
 626          * Get the next work request entry to find where to put the data.
 627          */
 628         if (qp->r_reuse_sge)
 629                 qp->r_reuse_sge = 0;
 630         else if (!ipath_get_rwqe(qp, 0)) {
 631                 /*
 632                  * Count VL15 packets dropped due to no receive buffer.
 633                  * Otherwise, count them as buffer overruns since usually,
 634                  * the HW will be able to receive packets even if there are
 635                  * no QPs with posted receive buffers.
 636                  */
 637                 if (qp->ibqp.qp_num == 0)
 638                         dev->n_vl15_dropped++;
 639                 else
 640                         dev->rcv_errors++;
 641                 goto bail;
 642         }
 643         /* Silently drop packets which are too big. */
 644         if (wc.byte_len > qp->r_len) {
 645                 qp->r_reuse_sge = 1;
 646                 dev->n_pkt_drops++;
 647                 goto bail;
 648         }
 649         if (has_grh) {
 650                 ipath_copy_sge(&qp->r_sge, &hdr->u.l.grh,
 651                                sizeof(struct ib_grh));
 652                 wc.wc_flags |= IB_WC_GRH;
 653         } else
 654                 ipath_skip_sge(&qp->r_sge, sizeof(struct ib_grh));
 655         ipath_copy_sge(&qp->r_sge, data,
 656                        wc.byte_len - sizeof(struct ib_grh));
 657         qp->r_wrid_valid = 0;
 658         wc.wr_id = qp->r_wr_id;
 659         wc.status = IB_WC_SUCCESS;
 660         wc.opcode = IB_WC_RECV;
 661         wc.vendor_err = 0;
 662         wc.qp = &qp->ibqp;
 663         wc.src_qp = src_qp;
 664         /* XXX do we know which pkey matched? Only needed for GSI. */
 665         wc.pkey_index = 0;
 666         wc.slid = be16_to_cpu(hdr->lrh[3]);
 667         wc.sl = (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF;
 668         dlid = be16_to_cpu(hdr->lrh[1]);
 669         /*
 670          * Save the LMC lower bits if the destination LID is a unicast LID.
 671          */
 672         wc.dlid_path_bits = dlid >= IPATH_MULTICAST_LID_BASE ? 0 :
 673                 dlid & ((1 << (dev->mkeyprot_resv_lmc & 7)) - 1);
 674         /* Signal completion event if the solicited bit is set. */
 675         ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
 676                        (ohdr->bth[0] &
 677                         __constant_cpu_to_be32(1 << 23)) != 0);
 678
 679 bail:;
 680 }