Add the rt linux 4.1.3-rt3 as base
[kvmfornfv.git] / kernel / drivers / infiniband / hw / qib / qib_ud.c
diff --git a/kernel/drivers/infiniband/hw/qib/qib_ud.c b/kernel/drivers/infiniband/hw/qib/qib_ud.c
new file mode 100644 (file)
index 0000000..26243b7
--- /dev/null
@@ -0,0 +1,590 @@
+/*
+ * Copyright (c) 2006, 2007, 2008, 2009 QLogic Corporation. All rights reserved.
+ * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include <rdma/ib_smi.h>
+
+#include "qib.h"
+#include "qib_mad.h"
+
+/**
+ * qib_ud_loopback - handle send on loopback QPs
+ * @sqp: the sending QP
+ * @swqe: the send work request
+ *
+ * This is called from qib_make_ud_req() to forward a WQE addressed
+ * to the same HCA.
+ * Note that the receive interrupt handler may be calling qib_ud_rcv()
+ * while this is being called.
+ */
+static void qib_ud_loopback(struct qib_qp *sqp, struct qib_swqe *swqe)
+{
+       struct qib_ibport *ibp = to_iport(sqp->ibqp.device, sqp->port_num);
+       struct qib_pportdata *ppd;
+       struct qib_qp *qp;
+       struct ib_ah_attr *ah_attr;
+       unsigned long flags;
+       struct qib_sge_state ssge;
+       struct qib_sge *sge;
+       struct ib_wc wc;
+       u32 length;
+       enum ib_qp_type sqptype, dqptype;
+
+       qp = qib_lookup_qpn(ibp, swqe->wr.wr.ud.remote_qpn);
+       if (!qp) {
+               ibp->n_pkt_drops++;
+               return;
+       }
+
+       sqptype = sqp->ibqp.qp_type == IB_QPT_GSI ?
+                       IB_QPT_UD : sqp->ibqp.qp_type;
+       dqptype = qp->ibqp.qp_type == IB_QPT_GSI ?
+                       IB_QPT_UD : qp->ibqp.qp_type;
+
+       if (dqptype != sqptype ||
+           !(ib_qib_state_ops[qp->state] & QIB_PROCESS_RECV_OK)) {
+               ibp->n_pkt_drops++;
+               goto drop;
+       }
+
+       ah_attr = &to_iah(swqe->wr.wr.ud.ah)->attr;
+       ppd = ppd_from_ibp(ibp);
+
+       if (qp->ibqp.qp_num > 1) {
+               u16 pkey1;
+               u16 pkey2;
+               u16 lid;
+
+               pkey1 = qib_get_pkey(ibp, sqp->s_pkey_index);
+               pkey2 = qib_get_pkey(ibp, qp->s_pkey_index);
+               if (unlikely(!qib_pkey_ok(pkey1, pkey2))) {
+                       lid = ppd->lid | (ah_attr->src_path_bits &
+                                         ((1 << ppd->lmc) - 1));
+                       qib_bad_pqkey(ibp, IB_NOTICE_TRAP_BAD_PKEY, pkey1,
+                                     ah_attr->sl,
+                                     sqp->ibqp.qp_num, qp->ibqp.qp_num,
+                                     cpu_to_be16(lid),
+                                     cpu_to_be16(ah_attr->dlid));
+                       goto drop;
+               }
+       }
+
+       /*
+        * Check that the qkey matches (except for QP0, see 9.6.1.4.1).
+        * Qkeys with the high order bit set mean use the
+        * qkey from the QP context instead of the WR (see 10.2.5).
+        */
+       if (qp->ibqp.qp_num) {
+               u32 qkey;
+
+               qkey = (int)swqe->wr.wr.ud.remote_qkey < 0 ?
+                       sqp->qkey : swqe->wr.wr.ud.remote_qkey;
+               if (unlikely(qkey != qp->qkey)) {
+                       u16 lid;
+
+                       lid = ppd->lid | (ah_attr->src_path_bits &
+                                         ((1 << ppd->lmc) - 1));
+                       qib_bad_pqkey(ibp, IB_NOTICE_TRAP_BAD_QKEY, qkey,
+                                     ah_attr->sl,
+                                     sqp->ibqp.qp_num, qp->ibqp.qp_num,
+                                     cpu_to_be16(lid),
+                                     cpu_to_be16(ah_attr->dlid));
+                       goto drop;
+               }
+       }
+
+       /*
+        * A GRH is expected to precede the data even if not
+        * present on the wire.
+        */
+       length = swqe->length;
+       memset(&wc, 0, sizeof(wc));
+       wc.byte_len = length + sizeof(struct ib_grh);
+
+       if (swqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
+               wc.wc_flags = IB_WC_WITH_IMM;
+               wc.ex.imm_data = swqe->wr.ex.imm_data;
+       }
+
+       spin_lock_irqsave(&qp->r_lock, flags);
+
+       /*
+        * Get the next work request entry to find where to put the data.
+        */
+       if (qp->r_flags & QIB_R_REUSE_SGE)
+               qp->r_flags &= ~QIB_R_REUSE_SGE;
+       else {
+               int ret;
+
+               ret = qib_get_rwqe(qp, 0);
+               if (ret < 0) {
+                       qib_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
+                       goto bail_unlock;
+               }
+               if (!ret) {
+                       if (qp->ibqp.qp_num == 0)
+                               ibp->n_vl15_dropped++;
+                       goto bail_unlock;
+               }
+       }
+       /* Silently drop packets which are too big. */
+       if (unlikely(wc.byte_len > qp->r_len)) {
+               qp->r_flags |= QIB_R_REUSE_SGE;
+               ibp->n_pkt_drops++;
+               goto bail_unlock;
+       }
+
+       if (ah_attr->ah_flags & IB_AH_GRH) {
+               qib_copy_sge(&qp->r_sge, &ah_attr->grh,
+                            sizeof(struct ib_grh), 1);
+               wc.wc_flags |= IB_WC_GRH;
+       } else
+               qib_skip_sge(&qp->r_sge, sizeof(struct ib_grh), 1);
+       ssge.sg_list = swqe->sg_list + 1;
+       ssge.sge = *swqe->sg_list;
+       ssge.num_sge = swqe->wr.num_sge;
+       sge = &ssge.sge;
+       while (length) {
+               u32 len = sge->length;
+
+               if (len > length)
+                       len = length;
+               if (len > sge->sge_length)
+                       len = sge->sge_length;
+               BUG_ON(len == 0);
+               qib_copy_sge(&qp->r_sge, sge->vaddr, len, 1);
+               sge->vaddr += len;
+               sge->length -= len;
+               sge->sge_length -= len;
+               if (sge->sge_length == 0) {
+                       if (--ssge.num_sge)
+                               *sge = *ssge.sg_list++;
+               } else if (sge->length == 0 && sge->mr->lkey) {
+                       if (++sge->n >= QIB_SEGSZ) {
+                               if (++sge->m >= sge->mr->mapsz)
+                                       break;
+                               sge->n = 0;
+                       }
+                       sge->vaddr =
+                               sge->mr->map[sge->m]->segs[sge->n].vaddr;
+                       sge->length =
+                               sge->mr->map[sge->m]->segs[sge->n].length;
+               }
+               length -= len;
+       }
+       qib_put_ss(&qp->r_sge);
+       if (!test_and_clear_bit(QIB_R_WRID_VALID, &qp->r_aflags))
+               goto bail_unlock;
+       wc.wr_id = qp->r_wr_id;
+       wc.status = IB_WC_SUCCESS;
+       wc.opcode = IB_WC_RECV;
+       wc.qp = &qp->ibqp;
+       wc.src_qp = sqp->ibqp.qp_num;
+       wc.pkey_index = qp->ibqp.qp_type == IB_QPT_GSI ?
+               swqe->wr.wr.ud.pkey_index : 0;
+       wc.slid = ppd->lid | (ah_attr->src_path_bits & ((1 << ppd->lmc) - 1));
+       wc.sl = ah_attr->sl;
+       wc.dlid_path_bits = ah_attr->dlid & ((1 << ppd->lmc) - 1);
+       wc.port_num = qp->port_num;
+       /* Signal completion event if the solicited bit is set. */
+       qib_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
+                    swqe->wr.send_flags & IB_SEND_SOLICITED);
+       ibp->n_loop_pkts++;
+bail_unlock:
+       spin_unlock_irqrestore(&qp->r_lock, flags);
+drop:
+       if (atomic_dec_and_test(&qp->refcount))
+               wake_up(&qp->wait);
+}
+
+/**
+ * qib_make_ud_req - construct a UD request packet
+ * @qp: the QP
+ *
+ * Return 1 if constructed; otherwise, return 0.
+ */
+int qib_make_ud_req(struct qib_qp *qp)
+{
+       struct qib_other_headers *ohdr;
+       struct ib_ah_attr *ah_attr;
+       struct qib_pportdata *ppd;
+       struct qib_ibport *ibp;
+       struct qib_swqe *wqe;
+       unsigned long flags;
+       u32 nwords;
+       u32 extra_bytes;
+       u32 bth0;
+       u16 lrh0;
+       u16 lid;
+       int ret = 0;
+       int next_cur;
+
+       spin_lock_irqsave(&qp->s_lock, flags);
+
+       if (!(ib_qib_state_ops[qp->state] & QIB_PROCESS_NEXT_SEND_OK)) {
+               if (!(ib_qib_state_ops[qp->state] & QIB_FLUSH_SEND))
+                       goto bail;
+               /* We are in the error state, flush the work request. */
+               if (qp->s_last == qp->s_head)
+                       goto bail;
+               /* If DMAs are in progress, we can't flush immediately. */
+               if (atomic_read(&qp->s_dma_busy)) {
+                       qp->s_flags |= QIB_S_WAIT_DMA;
+                       goto bail;
+               }
+               wqe = get_swqe_ptr(qp, qp->s_last);
+               qib_send_complete(qp, wqe, IB_WC_WR_FLUSH_ERR);
+               goto done;
+       }
+
+       if (qp->s_cur == qp->s_head)
+               goto bail;
+
+       wqe = get_swqe_ptr(qp, qp->s_cur);
+       next_cur = qp->s_cur + 1;
+       if (next_cur >= qp->s_size)
+               next_cur = 0;
+
+       /* Construct the header. */
+       ibp = to_iport(qp->ibqp.device, qp->port_num);
+       ppd = ppd_from_ibp(ibp);
+       ah_attr = &to_iah(wqe->wr.wr.ud.ah)->attr;
+       if (ah_attr->dlid >= QIB_MULTICAST_LID_BASE) {
+               if (ah_attr->dlid != QIB_PERMISSIVE_LID)
+                       this_cpu_inc(ibp->pmastats->n_multicast_xmit);
+               else
+                       this_cpu_inc(ibp->pmastats->n_unicast_xmit);
+       } else {
+               this_cpu_inc(ibp->pmastats->n_unicast_xmit);
+               lid = ah_attr->dlid & ~((1 << ppd->lmc) - 1);
+               if (unlikely(lid == ppd->lid)) {
+                       /*
+                        * If DMAs are in progress, we can't generate
+                        * a completion for the loopback packet since
+                        * it would be out of order.
+                        * XXX Instead of waiting, we could queue a
+                        * zero length descriptor so we get a callback.
+                        */
+                       if (atomic_read(&qp->s_dma_busy)) {
+                               qp->s_flags |= QIB_S_WAIT_DMA;
+                               goto bail;
+                       }
+                       qp->s_cur = next_cur;
+                       spin_unlock_irqrestore(&qp->s_lock, flags);
+                       qib_ud_loopback(qp, wqe);
+                       spin_lock_irqsave(&qp->s_lock, flags);
+                       qib_send_complete(qp, wqe, IB_WC_SUCCESS);
+                       goto done;
+               }
+       }
+
+       qp->s_cur = next_cur;
+       extra_bytes = -wqe->length & 3;
+       nwords = (wqe->length + extra_bytes) >> 2;
+
+       /* header size in 32-bit words LRH+BTH+DETH = (8+12+8)/4. */
+       qp->s_hdrwords = 7;
+       qp->s_cur_size = wqe->length;
+       qp->s_cur_sge = &qp->s_sge;
+       qp->s_srate = ah_attr->static_rate;
+       qp->s_wqe = wqe;
+       qp->s_sge.sge = wqe->sg_list[0];
+       qp->s_sge.sg_list = wqe->sg_list + 1;
+       qp->s_sge.num_sge = wqe->wr.num_sge;
+       qp->s_sge.total_len = wqe->length;
+
+       if (ah_attr->ah_flags & IB_AH_GRH) {
+               /* Header size in 32-bit words. */
+               qp->s_hdrwords += qib_make_grh(ibp, &qp->s_hdr->u.l.grh,
+                                              &ah_attr->grh,
+                                              qp->s_hdrwords, nwords);
+               lrh0 = QIB_LRH_GRH;
+               ohdr = &qp->s_hdr->u.l.oth;
+               /*
+                * Don't worry about sending to locally attached multicast
+                * QPs.  It is unspecified by the spec. what happens.
+                */
+       } else {
+               /* Header size in 32-bit words. */
+               lrh0 = QIB_LRH_BTH;
+               ohdr = &qp->s_hdr->u.oth;
+       }
+       if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
+               qp->s_hdrwords++;
+               ohdr->u.ud.imm_data = wqe->wr.ex.imm_data;
+               bth0 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE << 24;
+       } else
+               bth0 = IB_OPCODE_UD_SEND_ONLY << 24;
+       lrh0 |= ah_attr->sl << 4;
+       if (qp->ibqp.qp_type == IB_QPT_SMI)
+               lrh0 |= 0xF000; /* Set VL (see ch. 13.5.3.1) */
+       else
+               lrh0 |= ibp->sl_to_vl[ah_attr->sl] << 12;
+       qp->s_hdr->lrh[0] = cpu_to_be16(lrh0);
+       qp->s_hdr->lrh[1] = cpu_to_be16(ah_attr->dlid);  /* DEST LID */
+       qp->s_hdr->lrh[2] = cpu_to_be16(qp->s_hdrwords + nwords + SIZE_OF_CRC);
+       lid = ppd->lid;
+       if (lid) {
+               lid |= ah_attr->src_path_bits & ((1 << ppd->lmc) - 1);
+               qp->s_hdr->lrh[3] = cpu_to_be16(lid);
+       } else
+               qp->s_hdr->lrh[3] = IB_LID_PERMISSIVE;
+       if (wqe->wr.send_flags & IB_SEND_SOLICITED)
+               bth0 |= IB_BTH_SOLICITED;
+       bth0 |= extra_bytes << 20;
+       bth0 |= qp->ibqp.qp_type == IB_QPT_SMI ? QIB_DEFAULT_P_KEY :
+               qib_get_pkey(ibp, qp->ibqp.qp_type == IB_QPT_GSI ?
+                            wqe->wr.wr.ud.pkey_index : qp->s_pkey_index);
+       ohdr->bth[0] = cpu_to_be32(bth0);
+       /*
+        * Use the multicast QP if the destination LID is a multicast LID.
+        */
+       ohdr->bth[1] = ah_attr->dlid >= QIB_MULTICAST_LID_BASE &&
+               ah_attr->dlid != QIB_PERMISSIVE_LID ?
+               cpu_to_be32(QIB_MULTICAST_QPN) :
+               cpu_to_be32(wqe->wr.wr.ud.remote_qpn);
+       ohdr->bth[2] = cpu_to_be32(qp->s_next_psn++ & QIB_PSN_MASK);
+       /*
+        * Qkeys with the high order bit set mean use the
+        * qkey from the QP context instead of the WR (see 10.2.5).
+        */
+       ohdr->u.ud.deth[0] = cpu_to_be32((int)wqe->wr.wr.ud.remote_qkey < 0 ?
+                                        qp->qkey : wqe->wr.wr.ud.remote_qkey);
+       ohdr->u.ud.deth[1] = cpu_to_be32(qp->ibqp.qp_num);
+
+done:
+       ret = 1;
+       goto unlock;
+
+bail:
+       qp->s_flags &= ~QIB_S_BUSY;
+unlock:
+       spin_unlock_irqrestore(&qp->s_lock, flags);
+       return ret;
+}
+
+static unsigned qib_lookup_pkey(struct qib_ibport *ibp, u16 pkey)
+{
+       struct qib_pportdata *ppd = ppd_from_ibp(ibp);
+       struct qib_devdata *dd = ppd->dd;
+       unsigned ctxt = ppd->hw_pidx;
+       unsigned i;
+
+       pkey &= 0x7fff; /* remove limited/full membership bit */
+
+       for (i = 0; i < ARRAY_SIZE(dd->rcd[ctxt]->pkeys); ++i)
+               if ((dd->rcd[ctxt]->pkeys[i] & 0x7fff) == pkey)
+                       return i;
+
+       /*
+        * Should not get here, this means hardware failed to validate pkeys.
+        * Punt and return index 0.
+        */
+       return 0;
+}
+
+/**
+ * qib_ud_rcv - receive an incoming UD packet
+ * @ibp: the port the packet came in on
+ * @hdr: the packet header
+ * @has_grh: true if the packet has a GRH
+ * @data: the packet data
+ * @tlen: the packet length
+ * @qp: the QP the packet came on
+ *
+ * This is called from qib_qp_rcv() to process an incoming UD packet
+ * for the given QP.
+ * Called at interrupt level.
+ */
+void qib_ud_rcv(struct qib_ibport *ibp, struct qib_ib_header *hdr,
+               int has_grh, void *data, u32 tlen, struct qib_qp *qp)
+{
+       struct qib_other_headers *ohdr;
+       int opcode;
+       u32 hdrsize;
+       u32 pad;
+       struct ib_wc wc;
+       u32 qkey;
+       u32 src_qp;
+       u16 dlid;
+
+       /* Check for GRH */
+       if (!has_grh) {
+               ohdr = &hdr->u.oth;
+               hdrsize = 8 + 12 + 8;   /* LRH + BTH + DETH */
+       } else {
+               ohdr = &hdr->u.l.oth;
+               hdrsize = 8 + 40 + 12 + 8; /* LRH + GRH + BTH + DETH */
+       }
+       qkey = be32_to_cpu(ohdr->u.ud.deth[0]);
+       src_qp = be32_to_cpu(ohdr->u.ud.deth[1]) & QIB_QPN_MASK;
+
+       /*
+        * Get the number of bytes the message was padded by
+        * and drop incomplete packets.
+        */
+       pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
+       if (unlikely(tlen < (hdrsize + pad + 4)))
+               goto drop;
+
+       tlen -= hdrsize + pad + 4;
+
+       /*
+        * Check that the permissive LID is only used on QP0
+        * and the QKEY matches (see 9.6.1.4.1 and 9.6.1.5.1).
+        */
+       if (qp->ibqp.qp_num) {
+               if (unlikely(hdr->lrh[1] == IB_LID_PERMISSIVE ||
+                            hdr->lrh[3] == IB_LID_PERMISSIVE))
+                       goto drop;
+               if (qp->ibqp.qp_num > 1) {
+                       u16 pkey1, pkey2;
+
+                       pkey1 = be32_to_cpu(ohdr->bth[0]);
+                       pkey2 = qib_get_pkey(ibp, qp->s_pkey_index);
+                       if (unlikely(!qib_pkey_ok(pkey1, pkey2))) {
+                               qib_bad_pqkey(ibp, IB_NOTICE_TRAP_BAD_PKEY,
+                                             pkey1,
+                                             (be16_to_cpu(hdr->lrh[0]) >> 4) &
+                                               0xF,
+                                             src_qp, qp->ibqp.qp_num,
+                                             hdr->lrh[3], hdr->lrh[1]);
+                               return;
+                       }
+               }
+               if (unlikely(qkey != qp->qkey)) {
+                       qib_bad_pqkey(ibp, IB_NOTICE_TRAP_BAD_QKEY, qkey,
+                                     (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF,
+                                     src_qp, qp->ibqp.qp_num,
+                                     hdr->lrh[3], hdr->lrh[1]);
+                       return;
+               }
+               /* Drop invalid MAD packets (see 13.5.3.1). */
+               if (unlikely(qp->ibqp.qp_num == 1 &&
+                            (tlen != 256 ||
+                             (be16_to_cpu(hdr->lrh[0]) >> 12) == 15)))
+                       goto drop;
+       } else {
+               struct ib_smp *smp;
+
+               /* Drop invalid MAD packets (see 13.5.3.1). */
+               if (tlen != 256 || (be16_to_cpu(hdr->lrh[0]) >> 12) != 15)
+                       goto drop;
+               smp = (struct ib_smp *) data;
+               if ((hdr->lrh[1] == IB_LID_PERMISSIVE ||
+                    hdr->lrh[3] == IB_LID_PERMISSIVE) &&
+                   smp->mgmt_class != IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
+                       goto drop;
+       }
+
+       /*
+        * The opcode is in the low byte when its in network order
+        * (top byte when in host order).
+        */
+       opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
+       if (qp->ibqp.qp_num > 1 &&
+           opcode == IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE) {
+               wc.ex.imm_data = ohdr->u.ud.imm_data;
+               wc.wc_flags = IB_WC_WITH_IMM;
+               tlen -= sizeof(u32);
+       } else if (opcode == IB_OPCODE_UD_SEND_ONLY) {
+               wc.ex.imm_data = 0;
+               wc.wc_flags = 0;
+       } else
+               goto drop;
+
+       /*
+        * A GRH is expected to precede the data even if not
+        * present on the wire.
+        */
+       wc.byte_len = tlen + sizeof(struct ib_grh);
+
+       /*
+        * Get the next work request entry to find where to put the data.
+        */
+       if (qp->r_flags & QIB_R_REUSE_SGE)
+               qp->r_flags &= ~QIB_R_REUSE_SGE;
+       else {
+               int ret;
+
+               ret = qib_get_rwqe(qp, 0);
+               if (ret < 0) {
+                       qib_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
+                       return;
+               }
+               if (!ret) {
+                       if (qp->ibqp.qp_num == 0)
+                               ibp->n_vl15_dropped++;
+                       return;
+               }
+       }
+       /* Silently drop packets which are too big. */
+       if (unlikely(wc.byte_len > qp->r_len)) {
+               qp->r_flags |= QIB_R_REUSE_SGE;
+               goto drop;
+       }
+       if (has_grh) {
+               qib_copy_sge(&qp->r_sge, &hdr->u.l.grh,
+                            sizeof(struct ib_grh), 1);
+               wc.wc_flags |= IB_WC_GRH;
+       } else
+               qib_skip_sge(&qp->r_sge, sizeof(struct ib_grh), 1);
+       qib_copy_sge(&qp->r_sge, data, wc.byte_len - sizeof(struct ib_grh), 1);
+       qib_put_ss(&qp->r_sge);
+       if (!test_and_clear_bit(QIB_R_WRID_VALID, &qp->r_aflags))
+               return;
+       wc.wr_id = qp->r_wr_id;
+       wc.status = IB_WC_SUCCESS;
+       wc.opcode = IB_WC_RECV;
+       wc.vendor_err = 0;
+       wc.qp = &qp->ibqp;
+       wc.src_qp = src_qp;
+       wc.pkey_index = qp->ibqp.qp_type == IB_QPT_GSI ?
+               qib_lookup_pkey(ibp, be32_to_cpu(ohdr->bth[0])) : 0;
+       wc.slid = be16_to_cpu(hdr->lrh[3]);
+       wc.sl = (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF;
+       dlid = be16_to_cpu(hdr->lrh[1]);
+       /*
+        * Save the LMC lower bits if the destination LID is a unicast LID.
+        */
+       wc.dlid_path_bits = dlid >= QIB_MULTICAST_LID_BASE ? 0 :
+               dlid & ((1 << ppd_from_ibp(ibp)->lmc) - 1);
+       wc.port_num = qp->port_num;
+       /* Signal completion event if the solicited bit is set. */
+       qib_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
+                    (ohdr->bth[0] &
+                       cpu_to_be32(IB_BTH_SOLICITED)) != 0);
+       return;
+
+drop:
+       ibp->n_pkt_drops++;
+}