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                 BUG_ON(len == 0);
 235                 ipath_copy_sge(&rsge, sge->vaddr, len);
 236                 sge->vaddr += len;
 237                 sge->length -= len;
 238                 sge->sge_length -= len;
 239                 if (sge->sge_length == 0) {
 240                         if (--ss->num_sge)
 241                                 *sge = *ss->sg_list++;
 242                 } else if (sge->length == 0 && sge->mr != NULL) {
 243                         if (++sge->n >= IPATH_SEGSZ) {
 244                                 if (++sge->m >= sge->mr->mapsz)
 245                                         break;
 246                                 sge->n = 0;
 247                         }
 248                         sge->vaddr =
 249                                 sge->mr->map[sge->m]->segs[sge->n].vaddr;
 250                         sge->length =
 251                                 sge->mr->map[sge->m]->segs[sge->n].length;
 252                 }
 253                 length -= len;
 254         }
 255         wc->status = IB_WC_SUCCESS;
 256         wc->opcode = IB_WC_RECV;
 257         wc->vendor_err = 0;
 258         wc->qp = &qp->ibqp;
 259         wc->src_qp = sqp->ibqp.qp_num;
 260         /* XXX do we know which pkey matched? Only needed for GSI. */
 261         wc->pkey_index = 0;
 262         wc->slid = dev->dd->ipath_lid |
 263                 (ah_attr->src_path_bits &
 264                  ((1 << (dev->mkeyprot_resv_lmc & 7)) - 1));
 265         wc->sl = ah_attr->sl;
 266         wc->dlid_path_bits =
 267                 ah_attr->dlid & ((1 << (dev->mkeyprot_resv_lmc & 7)) - 1);
 268         /* Signal completion event if the solicited bit is set. */
 269         ipath_cq_enter(to_icq(qp->ibqp.recv_cq), wc,
 270                        wr->send_flags & IB_SEND_SOLICITED);
 271
 272 bail_sge:
 273         kfree(rsge.sg_list);
 274 done:
 275         if (atomic_dec_and_test(&qp->refcount))
 276                 wake_up(&qp->wait);
 277 }
 278
 279 /**
 280  * ipath_post_ud_send - post a UD send on QP
 281  * @qp: the QP
 282  * @wr: the work request
 283  *
 284  * Note that we actually send the data as it is posted instead of putting
 285  * the request into a ring buffer.  If we wanted to use a ring buffer,
 286  * we would need to save a reference to the destination address in the SWQE.
 287  */
 288 int ipath_post_ud_send(struct ipath_qp *qp, struct ib_send_wr *wr)
 289 {
 290         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
 291         struct ipath_other_headers *ohdr;
 292         struct ib_ah_attr *ah_attr;
 293         struct ipath_sge_state ss;
 294         struct ipath_sge *sg_list;
 295         struct ib_wc wc;
 296         u32 hwords;
 297         u32 nwords;
 298         u32 len;
 299         u32 extra_bytes;
 300         u32 bth0;
 301         u16 lrh0;
 302         u16 lid;
 303         int i;
 304         int ret;
 305
 306         if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK)) {
 307                 ret = 0;
 308                 goto bail;
 309         }
 310
 311         if (wr->wr.ud.ah->pd != qp->ibqp.pd) {
 312                 ret = -EPERM;
 313                 goto bail;
 314         }
 315
 316         /* IB spec says that num_sge == 0 is OK. */
 317         if (wr->num_sge > qp->s_max_sge) {
 318                 ret = -EINVAL;
 319                 goto bail;
 320         }
 321
 322         if (wr->num_sge > 1) {
 323                 sg_list = kmalloc((qp->s_max_sge - 1) * sizeof(*sg_list),
 324                                   GFP_ATOMIC);
 325                 if (!sg_list) {
 326                         ret = -ENOMEM;
 327                         goto bail;
 328                 }
 329         } else
 330                 sg_list = NULL;
 331
 332         /* Check the buffer to send. */
 333         ss.sg_list = sg_list;
 334         ss.sge.mr = NULL;
 335         ss.sge.vaddr = NULL;
 336         ss.sge.length = 0;
 337         ss.sge.sge_length = 0;
 338         ss.num_sge = 0;
 339         len = 0;
 340         for (i = 0; i < wr->num_sge; i++) {
 341                 /* Check LKEY */
 342                 if (to_ipd(qp->ibqp.pd)->user && wr->sg_list[i].lkey == 0) {
 343                         ret = -EINVAL;
 344                         goto bail;
 345                 }
 346
 347                 if (wr->sg_list[i].length == 0)
 348                         continue;
 349                 if (!ipath_lkey_ok(qp, ss.num_sge ?
 350                                    sg_list + ss.num_sge - 1 : &ss.sge,
 351                                    &wr->sg_list[i], 0)) {
 352                         ret = -EINVAL;
 353                         goto bail;
 354                 }
 355                 len += wr->sg_list[i].length;
 356                 ss.num_sge++;
 357         }
 358         /* Check for invalid packet size. */
 359         if (len > dev->dd->ipath_ibmtu) {
 360                 ret = -EINVAL;
 361                 goto bail;
 362         }
 363         extra_bytes = (4 - len) & 3;
 364         nwords = (len + extra_bytes) >> 2;
 365
 366         /* Construct the header. */
 367         ah_attr = &to_iah(wr->wr.ud.ah)->attr;
 368         if (ah_attr->dlid == 0) {
 369                 ret = -EINVAL;
 370                 goto bail;
 371         }
 372         if (ah_attr->dlid >= IPATH_MULTICAST_LID_BASE) {
 373                 if (ah_attr->dlid != IPATH_PERMISSIVE_LID)
 374                         dev->n_multicast_xmit++;
 375                 else
 376                         dev->n_unicast_xmit++;
 377         } else {
 378                 dev->n_unicast_xmit++;
 379                 lid = ah_attr->dlid &
 380                         ~((1 << (dev->mkeyprot_resv_lmc & 7)) - 1);
 381                 if (unlikely(lid == dev->dd->ipath_lid)) {
 382                         /*
 383                          * Pass in an uninitialized ib_wc to save stack
 384                          * space.
 385                          */
 386                         ipath_ud_loopback(qp, &ss, len, wr, &wc);
 387                         goto done;
 388                 }
 389         }
 390         if (ah_attr->ah_flags & IB_AH_GRH) {
 391                 /* Header size in 32-bit words. */
 392                 hwords = 17;
 393                 lrh0 = IPATH_LRH_GRH;
 394                 ohdr = &qp->s_hdr.u.l.oth;
 395                 qp->s_hdr.u.l.grh.version_tclass_flow =
 396                         cpu_to_be32((6 << 28) |
 397                                     (ah_attr->grh.traffic_class << 20) |
 398                                     ah_attr->grh.flow_label);
 399                 qp->s_hdr.u.l.grh.paylen =
 400                         cpu_to_be16(((wr->opcode ==
 401                                       IB_WR_SEND_WITH_IMM ? 6 : 5) +
 402                                      nwords + SIZE_OF_CRC) << 2);
 403                 /* next_hdr is defined by C8-7 in ch. 8.4.1 */
 404                 qp->s_hdr.u.l.grh.next_hdr = 0x1B;
 405                 qp->s_hdr.u.l.grh.hop_limit = ah_attr->grh.hop_limit;
 406                 /* The SGID is 32-bit aligned. */
 407                 qp->s_hdr.u.l.grh.sgid.global.subnet_prefix =
 408                         dev->gid_prefix;
 409                 qp->s_hdr.u.l.grh.sgid.global.interface_id =
 410                         dev->dd->ipath_guid;
 411                 qp->s_hdr.u.l.grh.dgid = ah_attr->grh.dgid;
 412                 /*
 413                  * Don't worry about sending to locally attached multicast
 414                  * QPs.  It is unspecified by the spec. what happens.
 415                  */
 416         } else {
 417                 /* Header size in 32-bit words. */
 418                 hwords = 7;
 419                 lrh0 = IPATH_LRH_BTH;
 420                 ohdr = &qp->s_hdr.u.oth;
 421         }
 422         if (wr->opcode == IB_WR_SEND_WITH_IMM) {
 423                 ohdr->u.ud.imm_data = wr->imm_data;
 424                 wc.imm_data = wr->imm_data;
 425                 hwords += 1;
 426                 bth0 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE << 24;
 427         } else if (wr->opcode == IB_WR_SEND) {
 428                 wc.imm_data = 0;
 429                 bth0 = IB_OPCODE_UD_SEND_ONLY << 24;
 430         } else {
 431                 ret = -EINVAL;
 432                 goto bail;
 433         }
 434         lrh0 |= ah_attr->sl << 4;
 435         if (qp->ibqp.qp_type == IB_QPT_SMI)
 436                 lrh0 |= 0xF000; /* Set VL (see ch. 13.5.3.1) */
 437         qp->s_hdr.lrh[0] = cpu_to_be16(lrh0);
 438         qp->s_hdr.lrh[1] = cpu_to_be16(ah_attr->dlid);  /* DEST LID */
 439         qp->s_hdr.lrh[2] = cpu_to_be16(hwords + nwords + SIZE_OF_CRC);
 440         lid = dev->dd->ipath_lid;
 441         if (lid) {
 442                 lid |= ah_attr->src_path_bits &
 443                         ((1 << (dev->mkeyprot_resv_lmc & 7)) - 1);
 444                 qp->s_hdr.lrh[3] = cpu_to_be16(lid);
 445         } else
 446                 qp->s_hdr.lrh[3] = IB_LID_PERMISSIVE;
 447         if (wr->send_flags & IB_SEND_SOLICITED)
 448                 bth0 |= 1 << 23;
 449         bth0 |= extra_bytes << 20;
 450         bth0 |= qp->ibqp.qp_type == IB_QPT_SMI ? IPATH_DEFAULT_P_KEY :
 451                 ipath_get_pkey(dev->dd, qp->s_pkey_index);
 452         ohdr->bth[0] = cpu_to_be32(bth0);
 453         /*
 454          * Use the multicast QP if the destination LID is a multicast LID.
 455          */
 456         ohdr->bth[1] = ah_attr->dlid >= IPATH_MULTICAST_LID_BASE &&
 457                 ah_attr->dlid != IPATH_PERMISSIVE_LID ?
 458                 __constant_cpu_to_be32(IPATH_MULTICAST_QPN) :
 459                 cpu_to_be32(wr->wr.ud.remote_qpn);
 460         /* XXX Could lose a PSN count but not worth locking */
 461         ohdr->bth[2] = cpu_to_be32(qp->s_next_psn++ & IPATH_PSN_MASK);
 462         /*
 463          * Qkeys with the high order bit set mean use the
 464          * qkey from the QP context instead of the WR (see 10.2.5).
 465          */
 466         ohdr->u.ud.deth[0] = cpu_to_be32((int)wr->wr.ud.remote_qkey < 0 ?
 467                                          qp->qkey : wr->wr.ud.remote_qkey);
 468         ohdr->u.ud.deth[1] = cpu_to_be32(qp->ibqp.qp_num);
 469         if (ipath_verbs_send(dev->dd, hwords, (u32 *) &qp->s_hdr,
 470                              len, &ss))
 471                 dev->n_no_piobuf++;
 472
 473 done:
 474         /* Queue the completion status entry. */
 475         if (!(qp->s_flags & IPATH_S_SIGNAL_REQ_WR) ||
 476             (wr->send_flags & IB_SEND_SIGNALED)) {
 477                 wc.wr_id = wr->wr_id;
 478                 wc.status = IB_WC_SUCCESS;
 479                 wc.vendor_err = 0;
 480                 wc.opcode = IB_WC_SEND;
 481                 wc.byte_len = len;
 482                 wc.qp = &qp->ibqp;
 483                 wc.src_qp = 0;
 484                 wc.wc_flags = 0;
 485                 /* XXX initialize other fields? */
 486                 ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 0);
 487         }
 488         kfree(sg_list);
 489
 490         ret = 0;
 491
 492 bail:
 493         return ret;
 494 }
 495
 496 /**
 497  * ipath_ud_rcv - receive an incoming UD packet
 498  * @dev: the device the packet came in on
 499  * @hdr: the packet header
 500  * @has_grh: true if the packet has a GRH
 501  * @data: the packet data
 502  * @tlen: the packet length
 503  * @qp: the QP the packet came on
 504  *
 505  * This is called from ipath_qp_rcv() to process an incoming UD packet
 506  * for the given QP.
 507  * Called at interrupt level.
 508  */
 509 void ipath_ud_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
 510                   int has_grh, void *data, u32 tlen, struct ipath_qp *qp)
 511 {
 512         struct ipath_other_headers *ohdr;
 513         int opcode;
 514         u32 hdrsize;
 515         u32 pad;
 516         struct ib_wc wc;
 517         u32 qkey;
 518         u32 src_qp;
 519         u16 dlid;
 520         int header_in_data;
 521
 522         /* Check for GRH */
 523         if (!has_grh) {
 524                 ohdr = &hdr->u.oth;
 525                 hdrsize = 8 + 12 + 8;   /* LRH + BTH + DETH */
 526                 qkey = be32_to_cpu(ohdr->u.ud.deth[0]);
 527                 src_qp = be32_to_cpu(ohdr->u.ud.deth[1]);
 528                 header_in_data = 0;
 529         } else {
 530                 ohdr = &hdr->u.l.oth;
 531                 hdrsize = 8 + 40 + 12 + 8; /* LRH + GRH + BTH + DETH */
 532                 /*
 533                  * The header with GRH is 68 bytes and the core driver sets
 534                  * the eager header buffer size to 56 bytes so the last 12
 535                  * bytes of the IB header is in the data buffer.
 536                  */
 537                 header_in_data = dev->dd->ipath_rcvhdrentsize == 16;
 538                 if (header_in_data) {
 539                         qkey = be32_to_cpu(((__be32 *) data)[1]);
 540                         src_qp = be32_to_cpu(((__be32 *) data)[2]);
 541                         data += 12;
 542                 } else {
 543                         qkey = be32_to_cpu(ohdr->u.ud.deth[0]);
 544                         src_qp = be32_to_cpu(ohdr->u.ud.deth[1]);
 545                 }
 546         }
 547         src_qp &= IPATH_QPN_MASK;
 548
 549         /*
 550          * Check that the permissive LID is only used on QP0
 551          * and the QKEY matches (see 9.6.1.4.1 and 9.6.1.5.1).
 552          */
 553         if (qp->ibqp.qp_num) {
 554                 if (unlikely(hdr->lrh[1] == IB_LID_PERMISSIVE ||
 555                              hdr->lrh[3] == IB_LID_PERMISSIVE)) {
 556                         dev->n_pkt_drops++;
 557                         goto bail;
 558                 }
 559                 if (unlikely(qkey != qp->qkey)) {
 560                         /* XXX OK to lose a count once in a while. */
 561                         dev->qkey_violations++;
 562                         dev->n_pkt_drops++;
 563                         goto bail;
 564                 }
 565         } else if (hdr->lrh[1] == IB_LID_PERMISSIVE ||
 566                    hdr->lrh[3] == IB_LID_PERMISSIVE) {
 567                 struct ib_smp *smp = (struct ib_smp *) data;
 568
 569                 if (smp->mgmt_class != IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
 570                         dev->n_pkt_drops++;
 571                         goto bail;
 572                 }
 573         }
 574
 575         /* Get the number of bytes the message was padded by. */
 576         pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
 577         if (unlikely(tlen < (hdrsize + pad + 4))) {
 578                 /* Drop incomplete packets. */
 579                 dev->n_pkt_drops++;
 580                 goto bail;
 581         }
 582         tlen -= hdrsize + pad + 4;
 583
 584         /* Drop invalid MAD packets (see 13.5.3.1). */
 585         if (unlikely((qp->ibqp.qp_num == 0 &&
 586                       (tlen != 256 ||
 587                        (be16_to_cpu(hdr->lrh[0]) >> 12) != 15)) ||
 588                      (qp->ibqp.qp_num == 1 &&
 589                       (tlen != 256 ||
 590                        (be16_to_cpu(hdr->lrh[0]) >> 12) == 15)))) {
 591                 dev->n_pkt_drops++;
 592                 goto bail;
 593         }
 594
 595         /*
 596          * A GRH is expected to preceed the data even if not
 597          * present on the wire.
 598          */
 599         wc.byte_len = tlen + sizeof(struct ib_grh);
 600
 601         /*
 602          * The opcode is in the low byte when its in network order
 603          * (top byte when in host order).
 604          */
 605         opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
 606         if (qp->ibqp.qp_num > 1 &&
 607             opcode == IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE) {
 608                 if (header_in_data) {
 609                         wc.imm_data = *(__be32 *) data;
 610                         data += sizeof(__be32);
 611                 } else
 612                         wc.imm_data = ohdr->u.ud.imm_data;
 613                 wc.wc_flags = IB_WC_WITH_IMM;
 614                 hdrsize += sizeof(u32);
 615         } else if (opcode == IB_OPCODE_UD_SEND_ONLY) {
 616                 wc.imm_data = 0;
 617                 wc.wc_flags = 0;
 618         } else {
 619                 dev->n_pkt_drops++;
 620                 goto bail;
 621         }
 622
 623         /*
 624          * Get the next work request entry to find where to put the data.
 625          */
 626         if (qp->r_reuse_sge)
 627                 qp->r_reuse_sge = 0;
 628         else if (!ipath_get_rwqe(qp, 0)) {
 629                 /*
 630                  * Count VL15 packets dropped due to no receive buffer.
 631                  * Otherwise, count them as buffer overruns since usually,
 632                  * the HW will be able to receive packets even if there are
 633                  * no QPs with posted receive buffers.
 634                  */
 635                 if (qp->ibqp.qp_num == 0)
 636                         dev->n_vl15_dropped++;
 637                 else
 638                         dev->rcv_errors++;
 639                 goto bail;
 640         }
 641         /* Silently drop packets which are too big. */
 642         if (wc.byte_len > qp->r_len) {
 643                 qp->r_reuse_sge = 1;
 644                 dev->n_pkt_drops++;
 645                 goto bail;
 646         }
 647         if (has_grh) {
 648                 ipath_copy_sge(&qp->r_sge, &hdr->u.l.grh,
 649                                sizeof(struct ib_grh));
 650                 wc.wc_flags |= IB_WC_GRH;
 651         } else
 652                 ipath_skip_sge(&qp->r_sge, sizeof(struct ib_grh));
 653         ipath_copy_sge(&qp->r_sge, data,
 654                        wc.byte_len - sizeof(struct ib_grh));
 655         qp->r_wrid_valid = 0;
 656         wc.wr_id = qp->r_wr_id;
 657         wc.status = IB_WC_SUCCESS;
 658         wc.opcode = IB_WC_RECV;
 659         wc.vendor_err = 0;
 660         wc.qp = &qp->ibqp;
 661         wc.src_qp = src_qp;
 662         /* XXX do we know which pkey matched? Only needed for GSI. */
 663         wc.pkey_index = 0;
 664         wc.slid = be16_to_cpu(hdr->lrh[3]);
 665         wc.sl = (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF;
 666         dlid = be16_to_cpu(hdr->lrh[1]);
 667         /*
 668          * Save the LMC lower bits if the destination LID is a unicast LID.
 669          */
 670         wc.dlid_path_bits = dlid >= IPATH_MULTICAST_LID_BASE ? 0 :
 671                 dlid & ((1 << (dev->mkeyprot_resv_lmc & 7)) - 1);
 672         /* Signal completion event if the solicited bit is set. */
 673         ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
 674                        (ohdr->bth[0] &
 675                         __constant_cpu_to_be32(1 << 23)) != 0);
 676
 677 bail:;
 678 }