IB/ipath: infiniband RC protocol support
authorBryan O'Sullivan <bos@pathscale.com>
Wed, 29 Mar 2006 23:23:35 +0000 (15:23 -0800)
committerRoland Dreier <rolandd@cisco.com>
Fri, 31 Mar 2006 21:14:20 +0000 (13:14 -0800)
This is an implementation of the Infiniband RC ("reliable connection")
protocol.

Signed-off-by: Bryan O'Sullivan <bos@pathscale.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
drivers/infiniband/hw/ipath/ipath_rc.c [new file with mode: 0644]

diff --git a/drivers/infiniband/hw/ipath/ipath_rc.c b/drivers/infiniband/hw/ipath/ipath_rc.c
new file mode 100644 (file)
index 0000000..a4055ca
--- /dev/null
@@ -0,0 +1,1857 @@
+/*
+ * 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 "ipath_verbs.h"
+#include "ips_common.h"
+
+/* cut down ridiculously long IB macro names */
+#define OP(x) IB_OPCODE_RC_##x
+
+/**
+ * ipath_init_restart- initialize the qp->s_sge after a restart
+ * @qp: the QP who's SGE we're restarting
+ * @wqe: the work queue to initialize the QP's SGE from
+ *
+ * The QP s_lock should be held.
+ */
+static void ipath_init_restart(struct ipath_qp *qp, struct ipath_swqe *wqe)
+{
+       struct ipath_ibdev *dev;
+       u32 len;
+
+       len = ((qp->s_psn - wqe->psn) & IPS_PSN_MASK) *
+               ib_mtu_enum_to_int(qp->path_mtu);
+       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;
+       ipath_skip_sge(&qp->s_sge, len);
+       qp->s_len = wqe->length - len;
+       dev = to_idev(qp->ibqp.device);
+       spin_lock(&dev->pending_lock);
+       if (qp->timerwait.next == LIST_POISON1)
+               list_add_tail(&qp->timerwait,
+                             &dev->pending[dev->pending_index]);
+       spin_unlock(&dev->pending_lock);
+}
+
+/**
+ * ipath_make_rc_ack - construct a response packet (ACK, NAK, or RDMA read)
+ * @qp: a pointer to the QP
+ * @ohdr: a pointer to the IB header being constructed
+ * @pmtu: the path MTU
+ *
+ * Return bth0 if constructed; otherwise, return 0.
+ * Note the QP s_lock must be held.
+ */
+static inline u32 ipath_make_rc_ack(struct ipath_qp *qp,
+                                   struct ipath_other_headers *ohdr,
+                                   u32 pmtu)
+{
+       struct ipath_sge_state *ss;
+       u32 hwords;
+       u32 len;
+       u32 bth0;
+
+       /* header size in 32-bit words LRH+BTH = (8+12)/4. */
+       hwords = 5;
+
+       /*
+        * Send a response.  Note that we are in the responder's
+        * side of the QP context.
+        */
+       switch (qp->s_ack_state) {
+       case OP(RDMA_READ_REQUEST):
+               ss = &qp->s_rdma_sge;
+               len = qp->s_rdma_len;
+               if (len > pmtu) {
+                       len = pmtu;
+                       qp->s_ack_state = OP(RDMA_READ_RESPONSE_FIRST);
+               }
+               else
+                       qp->s_ack_state = OP(RDMA_READ_RESPONSE_ONLY);
+               qp->s_rdma_len -= len;
+               bth0 = qp->s_ack_state << 24;
+               ohdr->u.aeth = ipath_compute_aeth(qp);
+               hwords++;
+               break;
+
+       case OP(RDMA_READ_RESPONSE_FIRST):
+               qp->s_ack_state = OP(RDMA_READ_RESPONSE_MIDDLE);
+               /* FALLTHROUGH */
+       case OP(RDMA_READ_RESPONSE_MIDDLE):
+               ss = &qp->s_rdma_sge;
+               len = qp->s_rdma_len;
+               if (len > pmtu)
+                       len = pmtu;
+               else {
+                       ohdr->u.aeth = ipath_compute_aeth(qp);
+                       hwords++;
+                       qp->s_ack_state = OP(RDMA_READ_RESPONSE_LAST);
+               }
+               qp->s_rdma_len -= len;
+               bth0 = qp->s_ack_state << 24;
+               break;
+
+       case OP(RDMA_READ_RESPONSE_LAST):
+       case OP(RDMA_READ_RESPONSE_ONLY):
+               /*
+                * We have to prevent new requests from changing
+                * the r_sge state while a ipath_verbs_send()
+                * is in progress.
+                * Changing r_state allows the receiver
+                * to continue processing new packets.
+                * We do it here now instead of above so
+                * that we are sure the packet was sent before
+                * changing the state.
+                */
+               qp->r_state = OP(RDMA_READ_RESPONSE_LAST);
+               qp->s_ack_state = OP(ACKNOWLEDGE);
+               return 0;
+
+       case OP(COMPARE_SWAP):
+       case OP(FETCH_ADD):
+               ss = NULL;
+               len = 0;
+               qp->r_state = OP(SEND_LAST);
+               qp->s_ack_state = OP(ACKNOWLEDGE);
+               bth0 = IB_OPCODE_ATOMIC_ACKNOWLEDGE << 24;
+               ohdr->u.at.aeth = ipath_compute_aeth(qp);
+               ohdr->u.at.atomic_ack_eth = cpu_to_be64(qp->s_ack_atomic);
+               hwords += sizeof(ohdr->u.at) / 4;
+               break;
+
+       default:
+               /* Send a regular ACK. */
+               ss = NULL;
+               len = 0;
+               qp->s_ack_state = OP(ACKNOWLEDGE);
+               bth0 = qp->s_ack_state << 24;
+               ohdr->u.aeth = ipath_compute_aeth(qp);
+               hwords++;
+       }
+       qp->s_hdrwords = hwords;
+       qp->s_cur_sge = ss;
+       qp->s_cur_size = len;
+
+       return bth0;
+}
+
+/**
+ * ipath_make_rc_req - construct a request packet (SEND, RDMA r/w, ATOMIC)
+ * @qp: a pointer to the QP
+ * @ohdr: a pointer to the IB header being constructed
+ * @pmtu: the path MTU
+ * @bth0p: pointer to the BTH opcode word
+ * @bth2p: pointer to the BTH PSN word
+ *
+ * Return 1 if constructed; otherwise, return 0.
+ * Note the QP s_lock must be held.
+ */
+static inline int ipath_make_rc_req(struct ipath_qp *qp,
+                                   struct ipath_other_headers *ohdr,
+                                   u32 pmtu, u32 *bth0p, u32 *bth2p)
+{
+       struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
+       struct ipath_sge_state *ss;
+       struct ipath_swqe *wqe;
+       u32 hwords;
+       u32 len;
+       u32 bth0;
+       u32 bth2;
+       char newreq;
+
+       if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK) ||
+           qp->s_rnr_timeout)
+               goto done;
+
+       /* header size in 32-bit words LRH+BTH = (8+12)/4. */
+       hwords = 5;
+       bth0 = 0;
+
+       /* Send a request. */
+       wqe = get_swqe_ptr(qp, qp->s_cur);
+       switch (qp->s_state) {
+       default:
+               /*
+                * Resend an old request or start a new one.
+                *
+                * We keep track of the current SWQE so that
+                * we don't reset the "furthest progress" state
+                * if we need to back up.
+                */
+               newreq = 0;
+               if (qp->s_cur == qp->s_tail) {
+                       /* Check if send work queue is empty. */
+                       if (qp->s_tail == qp->s_head)
+                               goto done;
+                       qp->s_psn = wqe->psn = qp->s_next_psn;
+                       newreq = 1;
+               }
+               /*
+                * Note that we have to be careful not to modify the
+                * original work request since we may need to resend
+                * it.
+                */
+               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_len = len = wqe->length;
+               ss = &qp->s_sge;
+               bth2 = 0;
+               switch (wqe->wr.opcode) {
+               case IB_WR_SEND:
+               case IB_WR_SEND_WITH_IMM:
+                       /* If no credit, return. */
+                       if (qp->s_lsn != (u32) -1 &&
+                           ipath_cmp24(wqe->ssn, qp->s_lsn + 1) > 0)
+                               goto done;
+                       wqe->lpsn = wqe->psn;
+                       if (len > pmtu) {
+                               wqe->lpsn += (len - 1) / pmtu;
+                               qp->s_state = OP(SEND_FIRST);
+                               len = pmtu;
+                               break;
+                       }
+                       if (wqe->wr.opcode == IB_WR_SEND)
+                               qp->s_state = OP(SEND_ONLY);
+                       else {
+                               qp->s_state = OP(SEND_ONLY_WITH_IMMEDIATE);
+                               /* Immediate data comes after the BTH */
+                               ohdr->u.imm_data = wqe->wr.imm_data;
+                               hwords += 1;
+                       }
+                       if (wqe->wr.send_flags & IB_SEND_SOLICITED)
+                               bth0 |= 1 << 23;
+                       bth2 = 1 << 31; /* Request ACK. */
+                       if (++qp->s_cur == qp->s_size)
+                               qp->s_cur = 0;
+                       break;
+
+               case IB_WR_RDMA_WRITE:
+                       if (newreq)
+                               qp->s_lsn++;
+                       /* FALLTHROUGH */
+               case IB_WR_RDMA_WRITE_WITH_IMM:
+                       /* If no credit, return. */
+                       if (qp->s_lsn != (u32) -1 &&
+                           ipath_cmp24(wqe->ssn, qp->s_lsn + 1) > 0)
+                               goto done;
+                       ohdr->u.rc.reth.vaddr =
+                               cpu_to_be64(wqe->wr.wr.rdma.remote_addr);
+                       ohdr->u.rc.reth.rkey =
+                               cpu_to_be32(wqe->wr.wr.rdma.rkey);
+                       ohdr->u.rc.reth.length = cpu_to_be32(len);
+                       hwords += sizeof(struct ib_reth) / 4;
+                       wqe->lpsn = wqe->psn;
+                       if (len > pmtu) {
+                               wqe->lpsn += (len - 1) / pmtu;
+                               qp->s_state = OP(RDMA_WRITE_FIRST);
+                               len = pmtu;
+                               break;
+                       }
+                       if (wqe->wr.opcode == IB_WR_RDMA_WRITE)
+                               qp->s_state = OP(RDMA_WRITE_ONLY);
+                       else {
+                               qp->s_state =
+                                       OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE);
+                               /* Immediate data comes
+                                * after RETH */
+                               ohdr->u.rc.imm_data = wqe->wr.imm_data;
+                               hwords += 1;
+                               if (wqe->wr.send_flags & IB_SEND_SOLICITED)
+                                       bth0 |= 1 << 23;
+                       }
+                       bth2 = 1 << 31; /* Request ACK. */
+                       if (++qp->s_cur == qp->s_size)
+                               qp->s_cur = 0;
+                       break;
+
+               case IB_WR_RDMA_READ:
+                       ohdr->u.rc.reth.vaddr =
+                               cpu_to_be64(wqe->wr.wr.rdma.remote_addr);
+                       ohdr->u.rc.reth.rkey =
+                               cpu_to_be32(wqe->wr.wr.rdma.rkey);
+                       ohdr->u.rc.reth.length = cpu_to_be32(len);
+                       qp->s_state = OP(RDMA_READ_REQUEST);
+                       hwords += sizeof(ohdr->u.rc.reth) / 4;
+                       if (newreq) {
+                               qp->s_lsn++;
+                               /*
+                                * Adjust s_next_psn to count the
+                                * expected number of responses.
+                                */
+                               if (len > pmtu)
+                                       qp->s_next_psn += (len - 1) / pmtu;
+                               wqe->lpsn = qp->s_next_psn++;
+                       }
+                       ss = NULL;
+                       len = 0;
+                       if (++qp->s_cur == qp->s_size)
+                               qp->s_cur = 0;
+                       break;
+
+               case IB_WR_ATOMIC_CMP_AND_SWP:
+               case IB_WR_ATOMIC_FETCH_AND_ADD:
+                       if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP)
+                               qp->s_state = OP(COMPARE_SWAP);
+                       else
+                               qp->s_state = OP(FETCH_ADD);
+                       ohdr->u.atomic_eth.vaddr = cpu_to_be64(
+                               wqe->wr.wr.atomic.remote_addr);
+                       ohdr->u.atomic_eth.rkey = cpu_to_be32(
+                               wqe->wr.wr.atomic.rkey);
+                       ohdr->u.atomic_eth.swap_data = cpu_to_be64(
+                               wqe->wr.wr.atomic.swap);
+                       ohdr->u.atomic_eth.compare_data = cpu_to_be64(
+                               wqe->wr.wr.atomic.compare_add);
+                       hwords += sizeof(struct ib_atomic_eth) / 4;
+                       if (newreq) {
+                               qp->s_lsn++;
+                               wqe->lpsn = wqe->psn;
+                       }
+                       if (++qp->s_cur == qp->s_size)
+                               qp->s_cur = 0;
+                       ss = NULL;
+                       len = 0;
+                       break;
+
+               default:
+                       goto done;
+               }
+               if (newreq) {
+                       qp->s_tail++;
+                       if (qp->s_tail >= qp->s_size)
+                               qp->s_tail = 0;
+               }
+               bth2 |= qp->s_psn++ & IPS_PSN_MASK;
+               if ((int)(qp->s_psn - qp->s_next_psn) > 0)
+                       qp->s_next_psn = qp->s_psn;
+               spin_lock(&dev->pending_lock);
+               if (qp->timerwait.next == LIST_POISON1)
+                       list_add_tail(&qp->timerwait,
+                                     &dev->pending[dev->pending_index]);
+               spin_unlock(&dev->pending_lock);
+               break;
+
+       case OP(RDMA_READ_RESPONSE_FIRST):
+               /*
+                * This case can only happen if a send is restarted.  See
+                * ipath_restart_rc().
+                */
+               ipath_init_restart(qp, wqe);
+               /* FALLTHROUGH */
+       case OP(SEND_FIRST):
+               qp->s_state = OP(SEND_MIDDLE);
+               /* FALLTHROUGH */
+       case OP(SEND_MIDDLE):
+               bth2 = qp->s_psn++ & IPS_PSN_MASK;
+               if ((int)(qp->s_psn - qp->s_next_psn) > 0)
+                       qp->s_next_psn = qp->s_psn;
+               ss = &qp->s_sge;
+               len = qp->s_len;
+               if (len > pmtu) {
+                       /*
+                        * Request an ACK every 1/2 MB to avoid retransmit
+                        * timeouts.
+                        */
+                       if (((wqe->length - len) % (512 * 1024)) == 0)
+                               bth2 |= 1 << 31;
+                       len = pmtu;
+                       break;
+               }
+               if (wqe->wr.opcode == IB_WR_SEND)
+                       qp->s_state = OP(SEND_LAST);
+               else {
+                       qp->s_state = OP(SEND_LAST_WITH_IMMEDIATE);
+                       /* Immediate data comes after the BTH */
+                       ohdr->u.imm_data = wqe->wr.imm_data;
+                       hwords += 1;
+               }
+               if (wqe->wr.send_flags & IB_SEND_SOLICITED)
+                       bth0 |= 1 << 23;
+               bth2 |= 1 << 31;        /* Request ACK. */
+               qp->s_cur++;
+               if (qp->s_cur >= qp->s_size)
+                       qp->s_cur = 0;
+               break;
+
+       case OP(RDMA_READ_RESPONSE_LAST):
+               /*
+                * This case can only happen if a RDMA write is restarted.
+                * See ipath_restart_rc().
+                */
+               ipath_init_restart(qp, wqe);
+               /* FALLTHROUGH */
+       case OP(RDMA_WRITE_FIRST):
+               qp->s_state = OP(RDMA_WRITE_MIDDLE);
+               /* FALLTHROUGH */
+       case OP(RDMA_WRITE_MIDDLE):
+               bth2 = qp->s_psn++ & IPS_PSN_MASK;
+               if ((int)(qp->s_psn - qp->s_next_psn) > 0)
+                       qp->s_next_psn = qp->s_psn;
+               ss = &qp->s_sge;
+               len = qp->s_len;
+               if (len > pmtu) {
+                       /*
+                        * Request an ACK every 1/2 MB to avoid retransmit
+                        * timeouts.
+                        */
+                       if (((wqe->length - len) % (512 * 1024)) == 0)
+                               bth2 |= 1 << 31;
+                       len = pmtu;
+                       break;
+               }
+               if (wqe->wr.opcode == IB_WR_RDMA_WRITE)
+                       qp->s_state = OP(RDMA_WRITE_LAST);
+               else {
+                       qp->s_state = OP(RDMA_WRITE_LAST_WITH_IMMEDIATE);
+                       /* Immediate data comes after the BTH */
+                       ohdr->u.imm_data = wqe->wr.imm_data;
+                       hwords += 1;
+                       if (wqe->wr.send_flags & IB_SEND_SOLICITED)
+                               bth0 |= 1 << 23;
+               }
+               bth2 |= 1 << 31;        /* Request ACK. */
+               qp->s_cur++;
+               if (qp->s_cur >= qp->s_size)
+                       qp->s_cur = 0;
+               break;
+
+       case OP(RDMA_READ_RESPONSE_MIDDLE):
+               /*
+                * This case can only happen if a RDMA read is restarted.
+                * See ipath_restart_rc().
+                */
+               ipath_init_restart(qp, wqe);
+               len = ((qp->s_psn - wqe->psn) & IPS_PSN_MASK) * pmtu;
+               ohdr->u.rc.reth.vaddr =
+                       cpu_to_be64(wqe->wr.wr.rdma.remote_addr + len);
+               ohdr->u.rc.reth.rkey =
+                       cpu_to_be32(wqe->wr.wr.rdma.rkey);
+               ohdr->u.rc.reth.length = cpu_to_be32(qp->s_len);
+               qp->s_state = OP(RDMA_READ_REQUEST);
+               hwords += sizeof(ohdr->u.rc.reth) / 4;
+               bth2 = qp->s_psn++ & IPS_PSN_MASK;
+               if ((int)(qp->s_psn - qp->s_next_psn) > 0)
+                       qp->s_next_psn = qp->s_psn;
+               ss = NULL;
+               len = 0;
+               qp->s_cur++;
+               if (qp->s_cur == qp->s_size)
+                       qp->s_cur = 0;
+               break;
+
+       case OP(RDMA_READ_REQUEST):
+       case OP(COMPARE_SWAP):
+       case OP(FETCH_ADD):
+               /*
+                * We shouldn't start anything new until this request is
+                * finished.  The ACK will handle rescheduling us.  XXX The
+                * number of outstanding ones is negotiated at connection
+                * setup time (see pg. 258,289)?  XXX Also, if we support
+                * multiple outstanding requests, we need to check the WQE
+                * IB_SEND_FENCE flag and not send a new request if a RDMA
+                * read or atomic is pending.
+                */
+               goto done;
+       }
+       qp->s_len -= len;
+       qp->s_hdrwords = hwords;
+       qp->s_cur_sge = ss;
+       qp->s_cur_size = len;
+       *bth0p = bth0 | (qp->s_state << 24);
+       *bth2p = bth2;
+       return 1;
+
+done:
+       return 0;
+}
+
+static inline void ipath_make_rc_grh(struct ipath_qp *qp,
+                                    struct ib_global_route *grh,
+                                    u32 nwords)
+{
+       struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
+
+       /* GRH header size in 32-bit words. */
+       qp->s_hdrwords += 10;
+       qp->s_hdr.u.l.grh.version_tclass_flow =
+               cpu_to_be32((6 << 28) |
+                           (grh->traffic_class << 20) |
+                           grh->flow_label);
+       qp->s_hdr.u.l.grh.paylen =
+               cpu_to_be16(((qp->s_hdrwords - 12) + nwords +
+                            SIZE_OF_CRC) << 2);
+       /* next_hdr is defined by C8-7 in ch. 8.4.1 */
+       qp->s_hdr.u.l.grh.next_hdr = 0x1B;
+       qp->s_hdr.u.l.grh.hop_limit = grh->hop_limit;
+       /* The SGID is 32-bit aligned. */
+       qp->s_hdr.u.l.grh.sgid.global.subnet_prefix = dev->gid_prefix;
+       qp->s_hdr.u.l.grh.sgid.global.interface_id =
+               ipath_layer_get_guid(dev->dd);
+       qp->s_hdr.u.l.grh.dgid = grh->dgid;
+}
+
+/**
+ * ipath_do_rc_send - perform a send on an RC QP
+ * @data: contains a pointer to the QP
+ *
+ * Process entries in the send work queue until credit or queue is
+ * exhausted.  Only allow one CPU to send a packet per QP (tasklet).
+ * Otherwise, after we drop the QP s_lock, two threads could send
+ * packets out of order.
+ */
+void ipath_do_rc_send(unsigned long data)
+{
+       struct ipath_qp *qp = (struct ipath_qp *)data;
+       struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
+       unsigned long flags;
+       u16 lrh0;
+       u32 nwords;
+       u32 extra_bytes;
+       u32 bth0;
+       u32 bth2;
+       u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
+       struct ipath_other_headers *ohdr;
+
+       if (test_and_set_bit(IPATH_S_BUSY, &qp->s_flags))
+               goto bail;
+
+       if (unlikely(qp->remote_ah_attr.dlid ==
+                    ipath_layer_get_lid(dev->dd))) {
+               struct ib_wc wc;
+
+               /*
+                * Pass in an uninitialized ib_wc to be consistent with
+                * other places where ipath_ruc_loopback() is called.
+                */
+               ipath_ruc_loopback(qp, &wc);
+               goto clear;
+       }
+
+       ohdr = &qp->s_hdr.u.oth;
+       if (qp->remote_ah_attr.ah_flags & IB_AH_GRH)
+               ohdr = &qp->s_hdr.u.l.oth;
+
+again:
+       /* Check for a constructed packet to be sent. */
+       if (qp->s_hdrwords != 0) {
+               /*
+                * If no PIO bufs are available, return.  An interrupt will
+                * call ipath_ib_piobufavail() when one is available.
+                */
+               _VERBS_INFO("h %u %p\n", qp->s_hdrwords, &qp->s_hdr);
+               _VERBS_INFO("d %u %p %u %p %u %u %u %u\n", qp->s_cur_size,
+                           qp->s_cur_sge->sg_list,
+                           qp->s_cur_sge->num_sge,
+                           qp->s_cur_sge->sge.vaddr,
+                           qp->s_cur_sge->sge.sge_length,
+                           qp->s_cur_sge->sge.length,
+                           qp->s_cur_sge->sge.m,
+                           qp->s_cur_sge->sge.n);
+               if (ipath_verbs_send(dev->dd, qp->s_hdrwords,
+                                    (u32 *) &qp->s_hdr, qp->s_cur_size,
+                                    qp->s_cur_sge)) {
+                       ipath_no_bufs_available(qp, dev);
+                       goto bail;
+               }
+               dev->n_unicast_xmit++;
+               /* Record that we sent the packet and s_hdr is empty. */
+               qp->s_hdrwords = 0;
+       }
+
+       /*
+        * The lock is needed to synchronize between setting
+        * qp->s_ack_state, resend timer, and post_send().
+        */
+       spin_lock_irqsave(&qp->s_lock, flags);
+
+       /* Sending responses has higher priority over sending requests. */
+       if (qp->s_ack_state != OP(ACKNOWLEDGE) &&
+           (bth0 = ipath_make_rc_ack(qp, ohdr, pmtu)) != 0)
+               bth2 = qp->s_ack_psn++ & IPS_PSN_MASK;
+       else if (!ipath_make_rc_req(qp, ohdr, pmtu, &bth0, &bth2))
+               goto done;
+
+       spin_unlock_irqrestore(&qp->s_lock, flags);
+
+       /* Construct the header. */
+       extra_bytes = (4 - qp->s_cur_size) & 3;
+       nwords = (qp->s_cur_size + extra_bytes) >> 2;
+       lrh0 = IPS_LRH_BTH;
+       if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
+               ipath_make_rc_grh(qp, &qp->remote_ah_attr.grh, nwords);
+               lrh0 = IPS_LRH_GRH;
+       }
+       lrh0 |= qp->remote_ah_attr.sl << 4;
+       qp->s_hdr.lrh[0] = cpu_to_be16(lrh0);
+       qp->s_hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
+       qp->s_hdr.lrh[2] = cpu_to_be16(qp->s_hdrwords + nwords +
+                                      SIZE_OF_CRC);
+       qp->s_hdr.lrh[3] = cpu_to_be16(ipath_layer_get_lid(dev->dd));
+       bth0 |= ipath_layer_get_pkey(dev->dd, qp->s_pkey_index);
+       bth0 |= extra_bytes << 20;
+       ohdr->bth[0] = cpu_to_be32(bth0);
+       ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
+       ohdr->bth[2] = cpu_to_be32(bth2);
+
+       /* Check for more work to do. */
+       goto again;
+
+done:
+       spin_unlock_irqrestore(&qp->s_lock, flags);
+clear:
+       clear_bit(IPATH_S_BUSY, &qp->s_flags);
+bail:
+       return;
+}
+
+static void send_rc_ack(struct ipath_qp *qp)
+{
+       struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
+       u16 lrh0;
+       u32 bth0;
+       struct ipath_other_headers *ohdr;
+
+       /* Construct the header. */
+       ohdr = &qp->s_hdr.u.oth;
+       lrh0 = IPS_LRH_BTH;
+       /* header size in 32-bit words LRH+BTH+AETH = (8+12+4)/4. */
+       qp->s_hdrwords = 6;
+       if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
+               ipath_make_rc_grh(qp, &qp->remote_ah_attr.grh, 0);
+               ohdr = &qp->s_hdr.u.l.oth;
+               lrh0 = IPS_LRH_GRH;
+       }
+       bth0 = ipath_layer_get_pkey(dev->dd, qp->s_pkey_index);
+       ohdr->u.aeth = ipath_compute_aeth(qp);
+       if (qp->s_ack_state >= OP(COMPARE_SWAP)) {
+               bth0 |= IB_OPCODE_ATOMIC_ACKNOWLEDGE << 24;
+               ohdr->u.at.atomic_ack_eth = cpu_to_be64(qp->s_ack_atomic);
+               qp->s_hdrwords += sizeof(ohdr->u.at.atomic_ack_eth) / 4;
+       }
+       else
+               bth0 |= OP(ACKNOWLEDGE) << 24;
+       lrh0 |= qp->remote_ah_attr.sl << 4;
+       qp->s_hdr.lrh[0] = cpu_to_be16(lrh0);
+       qp->s_hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
+       qp->s_hdr.lrh[2] = cpu_to_be16(qp->s_hdrwords + SIZE_OF_CRC);
+       qp->s_hdr.lrh[3] = cpu_to_be16(ipath_layer_get_lid(dev->dd));
+       ohdr->bth[0] = cpu_to_be32(bth0);
+       ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
+       ohdr->bth[2] = cpu_to_be32(qp->s_ack_psn & IPS_PSN_MASK);
+
+       /*
+        * If we can send the ACK, clear the ACK state.
+        */
+       if (ipath_verbs_send(dev->dd, qp->s_hdrwords, (u32 *) &qp->s_hdr,
+                            0, NULL) == 0) {
+               qp->s_ack_state = OP(ACKNOWLEDGE);
+               dev->n_rc_qacks++;
+               dev->n_unicast_xmit++;
+       }
+}
+
+/**
+ * ipath_restart_rc - back up requester to resend the last un-ACKed request
+ * @qp: the QP to restart
+ * @psn: packet sequence number for the request
+ * @wc: the work completion request
+ *
+ * The QP s_lock should be held.
+ */
+void ipath_restart_rc(struct ipath_qp *qp, u32 psn, struct ib_wc *wc)
+{
+       struct ipath_swqe *wqe = get_swqe_ptr(qp, qp->s_last);
+       struct ipath_ibdev *dev;
+       u32 n;
+
+       /*
+        * If there are no requests pending, we are done.
+        */
+       if (ipath_cmp24(psn, qp->s_next_psn) >= 0 ||
+           qp->s_last == qp->s_tail)
+               goto done;
+
+       if (qp->s_retry == 0) {
+               wc->wr_id = wqe->wr.wr_id;
+               wc->status = IB_WC_RETRY_EXC_ERR;
+               wc->opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
+               wc->vendor_err = 0;
+               wc->byte_len = 0;
+               wc->qp_num = qp->ibqp.qp_num;
+               wc->src_qp = qp->remote_qpn;
+               wc->pkey_index = 0;
+               wc->slid = qp->remote_ah_attr.dlid;
+               wc->sl = qp->remote_ah_attr.sl;
+               wc->dlid_path_bits = 0;
+               wc->port_num = 0;
+               ipath_sqerror_qp(qp, wc);
+               goto bail;
+       }
+       qp->s_retry--;
+
+       /*
+        * Remove the QP from the timeout queue.
+        * Note: it may already have been removed by ipath_ib_timer().
+        */
+       dev = to_idev(qp->ibqp.device);
+       spin_lock(&dev->pending_lock);
+       if (qp->timerwait.next != LIST_POISON1)
+               list_del(&qp->timerwait);
+       spin_unlock(&dev->pending_lock);
+
+       if (wqe->wr.opcode == IB_WR_RDMA_READ)
+               dev->n_rc_resends++;
+       else
+               dev->n_rc_resends += (int)qp->s_psn - (int)psn;
+
+       /*
+        * If we are starting the request from the beginning, let the normal
+        * send code handle initialization.
+        */
+       qp->s_cur = qp->s_last;
+       if (ipath_cmp24(psn, wqe->psn) <= 0) {
+               qp->s_state = OP(SEND_LAST);
+               qp->s_psn = wqe->psn;
+       } else {
+               n = qp->s_cur;
+               for (;;) {
+                       if (++n == qp->s_size)
+                               n = 0;
+                       if (n == qp->s_tail) {
+                               if (ipath_cmp24(psn, qp->s_next_psn) >= 0) {
+                                       qp->s_cur = n;
+                                       wqe = get_swqe_ptr(qp, n);
+                               }
+                               break;
+                       }
+                       wqe = get_swqe_ptr(qp, n);
+                       if (ipath_cmp24(psn, wqe->psn) < 0)
+                               break;
+                       qp->s_cur = n;
+               }
+               qp->s_psn = psn;
+
+               /*
+                * Reset the state to restart in the middle of a request.
+                * Don't change the s_sge, s_cur_sge, or s_cur_size.
+                * See ipath_do_rc_send().
+                */
+               switch (wqe->wr.opcode) {
+               case IB_WR_SEND:
+               case IB_WR_SEND_WITH_IMM:
+                       qp->s_state = OP(RDMA_READ_RESPONSE_FIRST);
+                       break;
+
+               case IB_WR_RDMA_WRITE:
+               case IB_WR_RDMA_WRITE_WITH_IMM:
+                       qp->s_state = OP(RDMA_READ_RESPONSE_LAST);
+                       break;
+
+               case IB_WR_RDMA_READ:
+                       qp->s_state =
+                               OP(RDMA_READ_RESPONSE_MIDDLE);
+                       break;
+
+               default:
+                       /*
+                        * This case shouldn't happen since its only
+                        * one PSN per req.
+                        */
+                       qp->s_state = OP(SEND_LAST);
+               }
+       }
+
+done:
+       tasklet_hi_schedule(&qp->s_task);
+
+bail:
+       return;
+}
+
+/**
+ * reset_psn - reset the QP state to send starting from PSN
+ * @qp: the QP
+ * @psn: the packet sequence number to restart at
+ *
+ * This is called from ipath_rc_rcv() to process an incoming RC ACK
+ * for the given QP.
+ * Called at interrupt level with the QP s_lock held.
+ */
+static void reset_psn(struct ipath_qp *qp, u32 psn)
+{
+       struct ipath_swqe *wqe;
+       u32 n;
+
+       n = qp->s_cur;
+       wqe = get_swqe_ptr(qp, n);
+       for (;;) {
+               if (++n == qp->s_size)
+                       n = 0;
+               if (n == qp->s_tail) {
+                       if (ipath_cmp24(psn, qp->s_next_psn) >= 0) {
+                               qp->s_cur = n;
+                               wqe = get_swqe_ptr(qp, n);
+                       }
+                       break;
+               }
+               wqe = get_swqe_ptr(qp, n);
+               if (ipath_cmp24(psn, wqe->psn) < 0)
+                       break;
+               qp->s_cur = n;
+       }
+       qp->s_psn = psn;
+
+       /*
+        * Set the state to restart in the middle of a
+        * request.  Don't change the s_sge, s_cur_sge, or
+        * s_cur_size.  See ipath_do_rc_send().
+        */
+       switch (wqe->wr.opcode) {
+       case IB_WR_SEND:
+       case IB_WR_SEND_WITH_IMM:
+               qp->s_state = OP(RDMA_READ_RESPONSE_FIRST);
+               break;
+
+       case IB_WR_RDMA_WRITE:
+       case IB_WR_RDMA_WRITE_WITH_IMM:
+               qp->s_state = OP(RDMA_READ_RESPONSE_LAST);
+               break;
+
+       case IB_WR_RDMA_READ:
+               qp->s_state = OP(RDMA_READ_RESPONSE_MIDDLE);
+               break;
+
+       default:
+               /*
+                * This case shouldn't happen since its only
+                * one PSN per req.
+                */
+               qp->s_state = OP(SEND_LAST);
+       }
+}
+
+/**
+ * do_rc_ack - process an incoming RC ACK
+ * @qp: the QP the ACK came in on
+ * @psn: the packet sequence number of the ACK
+ * @opcode: the opcode of the request that resulted in the ACK
+ *
+ * This is called from ipath_rc_rcv() to process an incoming RC ACK
+ * for the given QP.
+ * Called at interrupt level with the QP s_lock held.
+ * Returns 1 if OK, 0 if current operation should be aborted (NAK).
+ */
+static int do_rc_ack(struct ipath_qp *qp, u32 aeth, u32 psn, int opcode)
+{
+       struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
+       struct ib_wc wc;
+       struct ipath_swqe *wqe;
+       int ret = 0;
+
+       /*
+        * Remove the QP from the timeout queue (or RNR timeout queue).
+        * If ipath_ib_timer() has already removed it,
+        * it's OK since we hold the QP s_lock and ipath_restart_rc()
+        * just won't find anything to restart if we ACK everything.
+        */
+       spin_lock(&dev->pending_lock);
+       if (qp->timerwait.next != LIST_POISON1)
+               list_del(&qp->timerwait);
+       spin_unlock(&dev->pending_lock);
+
+       /*
+        * Note that NAKs implicitly ACK outstanding SEND and RDMA write
+        * requests and implicitly NAK RDMA read and atomic requests issued
+        * before the NAK'ed request.  The MSN won't include the NAK'ed
+        * request but will include an ACK'ed request(s).
+        */
+       wqe = get_swqe_ptr(qp, qp->s_last);
+
+       /* Nothing is pending to ACK/NAK. */
+       if (qp->s_last == qp->s_tail)
+               goto bail;
+
+       /*
+        * The MSN might be for a later WQE than the PSN indicates so
+        * only complete WQEs that the PSN finishes.
+        */
+       while (ipath_cmp24(psn, wqe->lpsn) >= 0) {
+               /* If we are ACKing a WQE, the MSN should be >= the SSN. */
+               if (ipath_cmp24(aeth, wqe->ssn) < 0)
+                       break;
+               /*
+                * If this request is a RDMA read or atomic, and the ACK is
+                * for a later operation, this ACK NAKs the RDMA read or
+                * atomic.  In other words, only a RDMA_READ_LAST or ONLY
+                * can ACK a RDMA read and likewise for atomic ops.  Note
+                * that the NAK case can only happen if relaxed ordering is
+                * used and requests are sent after an RDMA read or atomic
+                * is sent but before the response is received.
+                */
+               if ((wqe->wr.opcode == IB_WR_RDMA_READ &&
+                    opcode != OP(RDMA_READ_RESPONSE_LAST)) ||
+                   ((wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
+                     wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) &&
+                    (opcode != OP(ATOMIC_ACKNOWLEDGE) ||
+                     ipath_cmp24(wqe->psn, psn) != 0))) {
+                       /*
+                        * The last valid PSN seen is the previous
+                        * request's.
+                        */
+                       qp->s_last_psn = wqe->psn - 1;
+                       /* Retry this request. */
+                       ipath_restart_rc(qp, wqe->psn, &wc);
+                       /*
+                        * No need to process the ACK/NAK since we are
+                        * restarting an earlier request.
+                        */
+                       goto bail;
+               }
+               /* Post a send completion queue entry if requested. */
+               if (!test_bit(IPATH_S_SIGNAL_REQ_WR, &qp->s_flags) ||
+                   (wqe->wr.send_flags & IB_SEND_SIGNALED)) {
+                       wc.wr_id = wqe->wr.wr_id;
+                       wc.status = IB_WC_SUCCESS;
+                       wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
+                       wc.vendor_err = 0;
+                       wc.byte_len = wqe->length;
+                       wc.qp_num = qp->ibqp.qp_num;
+                       wc.src_qp = qp->remote_qpn;
+                       wc.pkey_index = 0;
+                       wc.slid = qp->remote_ah_attr.dlid;
+                       wc.sl = qp->remote_ah_attr.sl;
+                       wc.dlid_path_bits = 0;
+                       wc.port_num = 0;
+                       ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 0);
+               }
+               qp->s_retry = qp->s_retry_cnt;
+               /*
+                * If we are completing a request which is in the process of
+                * being resent, we can stop resending it since we know the
+                * responder has already seen it.
+                */
+               if (qp->s_last == qp->s_cur) {
+                       if (++qp->s_cur >= qp->s_size)
+                               qp->s_cur = 0;
+                       wqe = get_swqe_ptr(qp, qp->s_cur);
+                       qp->s_state = OP(SEND_LAST);
+                       qp->s_psn = wqe->psn;
+               }
+               if (++qp->s_last >= qp->s_size)
+                       qp->s_last = 0;
+               wqe = get_swqe_ptr(qp, qp->s_last);
+               if (qp->s_last == qp->s_tail)
+                       break;
+       }
+
+       switch (aeth >> 29) {
+       case 0:         /* ACK */
+               dev->n_rc_acks++;
+               /* If this is a partial ACK, reset the retransmit timer. */
+               if (qp->s_last != qp->s_tail) {
+                       spin_lock(&dev->pending_lock);
+                       list_add_tail(&qp->timerwait,
+                                     &dev->pending[dev->pending_index]);
+                       spin_unlock(&dev->pending_lock);
+               }
+               ipath_get_credit(qp, aeth);
+               qp->s_rnr_retry = qp->s_rnr_retry_cnt;
+               qp->s_retry = qp->s_retry_cnt;
+               qp->s_last_psn = psn;
+               ret = 1;
+               goto bail;
+
+       case 1:         /* RNR NAK */
+               dev->n_rnr_naks++;
+               if (qp->s_rnr_retry == 0) {
+                       if (qp->s_last == qp->s_tail)
+                               goto bail;
+
+                       wc.status = IB_WC_RNR_RETRY_EXC_ERR;
+                       goto class_b;
+               }
+               if (qp->s_rnr_retry_cnt < 7)
+                       qp->s_rnr_retry--;
+               if (qp->s_last == qp->s_tail)
+                       goto bail;
+
+               /* The last valid PSN seen is the previous request's. */
+               qp->s_last_psn = wqe->psn - 1;
+
+               dev->n_rc_resends += (int)qp->s_psn - (int)psn;
+
+               /*
+                * If we are starting the request from the beginning, let
+                * the normal send code handle initialization.
+                */
+               qp->s_cur = qp->s_last;
+               wqe = get_swqe_ptr(qp, qp->s_cur);
+               if (ipath_cmp24(psn, wqe->psn) <= 0) {
+                       qp->s_state = OP(SEND_LAST);
+                       qp->s_psn = wqe->psn;
+               } else
+                       reset_psn(qp, psn);
+
+               qp->s_rnr_timeout =
+                       ib_ipath_rnr_table[(aeth >> IPS_AETH_CREDIT_SHIFT) &
+                                          IPS_AETH_CREDIT_MASK];
+               ipath_insert_rnr_queue(qp);
+               goto bail;
+
+       case 3:         /* NAK */
+               /* The last valid PSN seen is the previous request's. */
+               if (qp->s_last != qp->s_tail)
+                       qp->s_last_psn = wqe->psn - 1;
+               switch ((aeth >> IPS_AETH_CREDIT_SHIFT) &
+                       IPS_AETH_CREDIT_MASK) {
+               case 0: /* PSN sequence error */
+                       dev->n_seq_naks++;
+                       /*
+                        * Back up to the responder's expected PSN.  XXX
+                        * Note that we might get a NAK in the middle of an
+                        * RDMA READ response which terminates the RDMA
+                        * READ.
+                        */
+                       if (qp->s_last == qp->s_tail)
+                               break;
+
+                       if (ipath_cmp24(psn, wqe->psn) < 0)
+                               break;
+
+                       /* Retry the request. */
+                       ipath_restart_rc(qp, psn, &wc);
+                       break;
+
+               case 1: /* Invalid Request */
+                       wc.status = IB_WC_REM_INV_REQ_ERR;
+                       dev->n_other_naks++;
+                       goto class_b;
+
+               case 2: /* Remote Access Error */
+                       wc.status = IB_WC_REM_ACCESS_ERR;
+                       dev->n_other_naks++;
+                       goto class_b;
+
+               case 3: /* Remote Operation Error */
+                       wc.status = IB_WC_REM_OP_ERR;
+                       dev->n_other_naks++;
+               class_b:
+                       wc.wr_id = wqe->wr.wr_id;
+                       wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
+                       wc.vendor_err = 0;
+                       wc.byte_len = 0;
+                       wc.qp_num = qp->ibqp.qp_num;
+                       wc.src_qp = qp->remote_qpn;
+                       wc.pkey_index = 0;
+                       wc.slid = qp->remote_ah_attr.dlid;
+                       wc.sl = qp->remote_ah_attr.sl;
+                       wc.dlid_path_bits = 0;
+                       wc.port_num = 0;
+                       ipath_sqerror_qp(qp, &wc);
+                       break;
+
+               default:
+                       /* Ignore other reserved NAK error codes */
+                       goto reserved;
+               }
+               qp->s_rnr_retry = qp->s_rnr_retry_cnt;
+               goto bail;
+
+       default:                /* 2: reserved */
+       reserved:
+               /* Ignore reserved NAK codes. */
+               goto bail;
+       }
+
+bail:
+       return ret;
+}
+
+/**
+ * ipath_rc_rcv_resp - process an incoming RC response packet
+ * @dev: the device this packet came in on
+ * @ohdr: the other headers for this packet
+ * @data: the packet data
+ * @tlen: the packet length
+ * @qp: the QP for this packet
+ * @opcode: the opcode for this packet
+ * @psn: the packet sequence number for this packet
+ * @hdrsize: the header length
+ * @pmtu: the path MTU
+ * @header_in_data: true if part of the header data is in the data buffer
+ *
+ * This is called from ipath_rc_rcv() to process an incoming RC response
+ * packet for the given QP.
+ * Called at interrupt level.
+ */
+static inline void ipath_rc_rcv_resp(struct ipath_ibdev *dev,
+                                    struct ipath_other_headers *ohdr,
+                                    void *data, u32 tlen,
+                                    struct ipath_qp *qp,
+                                    u32 opcode,
+                                    u32 psn, u32 hdrsize, u32 pmtu,
+                                    int header_in_data)
+{
+       unsigned long flags;
+       struct ib_wc wc;
+       int diff;
+       u32 pad;
+       u32 aeth;
+
+       spin_lock_irqsave(&qp->s_lock, flags);
+
+       /* Ignore invalid responses. */
+       if (ipath_cmp24(psn, qp->s_next_psn) >= 0)
+               goto ack_done;
+
+       /* Ignore duplicate responses. */
+       diff = ipath_cmp24(psn, qp->s_last_psn);
+       if (unlikely(diff <= 0)) {
+               /* Update credits for "ghost" ACKs */
+               if (diff == 0 && opcode == OP(ACKNOWLEDGE)) {
+                       if (!header_in_data)
+                               aeth = be32_to_cpu(ohdr->u.aeth);
+                       else {
+                               aeth = be32_to_cpu(((__be32 *) data)[0]);
+                               data += sizeof(__be32);
+                       }
+                       if ((aeth >> 29) == 0)
+                               ipath_get_credit(qp, aeth);
+               }
+               goto ack_done;
+       }
+
+       switch (opcode) {
+       case OP(ACKNOWLEDGE):
+       case OP(ATOMIC_ACKNOWLEDGE):
+       case OP(RDMA_READ_RESPONSE_FIRST):
+               if (!header_in_data)
+                       aeth = be32_to_cpu(ohdr->u.aeth);
+               else {
+                       aeth = be32_to_cpu(((__be32 *) data)[0]);
+                       data += sizeof(__be32);
+               }
+               if (opcode == OP(ATOMIC_ACKNOWLEDGE))
+                       *(u64 *) qp->s_sge.sge.vaddr = *(u64 *) data;
+               if (!do_rc_ack(qp, aeth, psn, opcode) ||
+                   opcode != OP(RDMA_READ_RESPONSE_FIRST))
+                       goto ack_done;
+               hdrsize += 4;
+               /*
+                * do_rc_ack() has already checked the PSN so skip
+                * the sequence check.
+                */
+               goto rdma_read;
+
+       case OP(RDMA_READ_RESPONSE_MIDDLE):
+               /* no AETH, no ACK */
+               if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) {
+                       dev->n_rdma_seq++;
+                       ipath_restart_rc(qp, qp->s_last_psn + 1, &wc);
+                       goto ack_done;
+               }
+       rdma_read:
+       if (unlikely(qp->s_state != OP(RDMA_READ_REQUEST)))
+               goto ack_done;
+       if (unlikely(tlen != (hdrsize + pmtu + 4)))
+               goto ack_done;
+       if (unlikely(pmtu >= qp->s_len))
+               goto ack_done;
+       /* We got a response so update the timeout. */
+       if (unlikely(qp->s_last == qp->s_tail ||
+                    get_swqe_ptr(qp, qp->s_last)->wr.opcode !=
+                    IB_WR_RDMA_READ))
+               goto ack_done;
+       spin_lock(&dev->pending_lock);
+       if (qp->s_rnr_timeout == 0 &&
+           qp->timerwait.next != LIST_POISON1)
+               list_move_tail(&qp->timerwait,
+                              &dev->pending[dev->pending_index]);
+       spin_unlock(&dev->pending_lock);
+       /*
+        * Update the RDMA receive state but do the copy w/o holding the
+        * locks and blocking interrupts.  XXX Yet another place that
+        * affects relaxed RDMA order since we don't want s_sge modified.
+        */
+       qp->s_len -= pmtu;
+       qp->s_last_psn = psn;
+       spin_unlock_irqrestore(&qp->s_lock, flags);
+       ipath_copy_sge(&qp->s_sge, data, pmtu);
+       goto bail;
+
+       case OP(RDMA_READ_RESPONSE_LAST):
+               /* ACKs READ req. */
+               if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) {
+                       dev->n_rdma_seq++;
+                       ipath_restart_rc(qp, qp->s_last_psn + 1, &wc);
+                       goto ack_done;
+               }
+               /* FALLTHROUGH */
+       case OP(RDMA_READ_RESPONSE_ONLY):
+               if (unlikely(qp->s_state != OP(RDMA_READ_REQUEST)))
+                       goto ack_done;
+               /*
+                * Get the number of bytes the message was padded by.
+                */
+               pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
+               /*
+                * Check that the data size is >= 1 && <= pmtu.
+                * Remember to account for the AETH header (4) and
+                * ICRC (4).
+                */
+               if (unlikely(tlen <= (hdrsize + pad + 8))) {
+                       /*
+                        * XXX Need to generate an error CQ
+                        * entry.
+                        */
+                       goto ack_done;
+               }
+               tlen -= hdrsize + pad + 8;
+               if (unlikely(tlen != qp->s_len)) {
+                       /*
+                        * XXX Need to generate an error CQ
+                        * entry.
+                        */
+                       goto ack_done;
+               }
+               if (!header_in_data)
+                       aeth = be32_to_cpu(ohdr->u.aeth);
+               else {
+                       aeth = be32_to_cpu(((__be32 *) data)[0]);
+                       data += sizeof(__be32);
+               }
+               ipath_copy_sge(&qp->s_sge, data, tlen);
+               if (do_rc_ack(qp, aeth, psn, OP(RDMA_READ_RESPONSE_LAST))) {
+                       /*
+                        * Change the state so we contimue
+                        * processing new requests.
+                        */
+                       qp->s_state = OP(SEND_LAST);
+               }
+               goto ack_done;
+       }
+
+ack_done:
+       spin_unlock_irqrestore(&qp->s_lock, flags);
+bail:
+       return;
+}
+
+/**
+ * ipath_rc_rcv_error - process an incoming duplicate or error RC packet
+ * @dev: the device this packet came in on
+ * @ohdr: the other headers for this packet
+ * @data: the packet data
+ * @qp: the QP for this packet
+ * @opcode: the opcode for this packet
+ * @psn: the packet sequence number for this packet
+ * @diff: the difference between the PSN and the expected PSN
+ * @header_in_data: true if part of the header data is in the data buffer
+ *
+ * This is called from ipath_rc_rcv() to process an unexpected
+ * incoming RC packet for the given QP.
+ * Called at interrupt level.
+ * Return 1 if no more processing is needed; otherwise return 0 to
+ * schedule a response to be sent and the s_lock unlocked.
+ */
+static inline int ipath_rc_rcv_error(struct ipath_ibdev *dev,
+                                    struct ipath_other_headers *ohdr,
+                                    void *data,
+                                    struct ipath_qp *qp,
+                                    u32 opcode,
+                                    u32 psn,
+                                    int diff,
+                                    int header_in_data)
+{
+       struct ib_reth *reth;
+
+       if (diff > 0) {
+               /*
+                * Packet sequence error.
+                * A NAK will ACK earlier sends and RDMA writes.
+                * Don't queue the NAK if a RDMA read, atomic, or
+                * NAK is pending though.
+                */
+               spin_lock(&qp->s_lock);
+               if ((qp->s_ack_state >= OP(RDMA_READ_REQUEST) &&
+                    qp->s_ack_state != IB_OPCODE_ACKNOWLEDGE) ||
+                   qp->s_nak_state != 0) {
+                       spin_unlock(&qp->s_lock);
+                       goto done;
+               }
+               qp->s_ack_state = OP(SEND_ONLY);
+               qp->s_nak_state = IB_NAK_PSN_ERROR;
+               /* Use the expected PSN. */
+               qp->s_ack_psn = qp->r_psn;
+               goto resched;
+       }
+
+       /*
+        * Handle a duplicate request.  Don't re-execute SEND, RDMA
+        * write or atomic op.  Don't NAK errors, just silently drop
+        * the duplicate request.  Note that r_sge, r_len, and
+        * r_rcv_len may be in use so don't modify them.
+        *
+        * We are supposed to ACK the earliest duplicate PSN but we
+        * can coalesce an outstanding duplicate ACK.  We have to
+        * send the earliest so that RDMA reads can be restarted at
+        * the requester's expected PSN.
+        */
+       spin_lock(&qp->s_lock);
+       if (qp->s_ack_state != IB_OPCODE_ACKNOWLEDGE &&
+           ipath_cmp24(psn, qp->s_ack_psn) >= 0) {
+               if (qp->s_ack_state < IB_OPCODE_RDMA_READ_REQUEST)
+                       qp->s_ack_psn = psn;
+               spin_unlock(&qp->s_lock);
+               goto done;
+       }
+       switch (opcode) {
+       case OP(RDMA_READ_REQUEST):
+               /*
+                * We have to be careful to not change s_rdma_sge
+                * while ipath_do_rc_send() is using it and not
+                * holding the s_lock.
+                */
+               if (qp->s_ack_state != OP(ACKNOWLEDGE) &&
+                   qp->s_ack_state >= IB_OPCODE_RDMA_READ_REQUEST) {
+                       spin_unlock(&qp->s_lock);
+                       dev->n_rdma_dup_busy++;
+                       goto done;
+               }
+               /* RETH comes after BTH */
+               if (!header_in_data)
+                       reth = &ohdr->u.rc.reth;
+               else {
+                       reth = (struct ib_reth *)data;
+                       data += sizeof(*reth);
+               }
+               qp->s_rdma_len = be32_to_cpu(reth->length);
+               if (qp->s_rdma_len != 0) {
+                       u32 rkey = be32_to_cpu(reth->rkey);
+                       u64 vaddr = be64_to_cpu(reth->vaddr);
+                       int ok;
+
+                       /*
+                        * Address range must be a subset of the original
+                        * request and start on pmtu boundaries.
+                        */
+                       ok = ipath_rkey_ok(dev, &qp->s_rdma_sge,
+                                          qp->s_rdma_len, vaddr, rkey,
+                                          IB_ACCESS_REMOTE_READ);
+                       if (unlikely(!ok))
+                               goto done;
+               } else {
+                       qp->s_rdma_sge.sg_list = NULL;
+                       qp->s_rdma_sge.num_sge = 0;
+                       qp->s_rdma_sge.sge.mr = NULL;
+                       qp->s_rdma_sge.sge.vaddr = NULL;
+                       qp->s_rdma_sge.sge.length = 0;
+                       qp->s_rdma_sge.sge.sge_length = 0;
+               }
+               break;
+
+       case OP(COMPARE_SWAP):
+       case OP(FETCH_ADD):
+               /*
+                * Check for the PSN of the last atomic operations
+                * performed and resend the result if found.
+                */
+               if ((psn & IPS_PSN_MASK) != qp->r_atomic_psn) {
+                       spin_unlock(&qp->s_lock);
+                       goto done;
+               }
+               qp->s_ack_atomic = qp->r_atomic_data;
+               break;
+       }
+       qp->s_ack_state = opcode;
+       qp->s_nak_state = 0;
+       qp->s_ack_psn = psn;
+resched:
+       return 0;
+
+done:
+       return 1;
+}
+
+/**
+ * ipath_rc_rcv - process an incoming RC packet
+ * @dev: the device this packet came in on
+ * @hdr: the header of this packet
+ * @has_grh: true if the header has a GRH
+ * @data: the packet data
+ * @tlen: the packet length
+ * @qp: the QP for this packet
+ *
+ * This is called from ipath_qp_rcv() to process an incoming RC packet
+ * for the given QP.
+ * Called at interrupt level.
+ */
+void ipath_rc_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
+                 int has_grh, void *data, u32 tlen, struct ipath_qp *qp)
+{
+       struct ipath_other_headers *ohdr;
+       u32 opcode;
+       u32 hdrsize;
+       u32 psn;
+       u32 pad;
+       unsigned long flags;
+       struct ib_wc wc;
+       u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
+       int diff;
+       struct ib_reth *reth;
+       int header_in_data;
+
+       /* Check for GRH */
+       if (!has_grh) {
+               ohdr = &hdr->u.oth;
+               hdrsize = 8 + 12;       /* LRH + BTH */
+               psn = be32_to_cpu(ohdr->bth[2]);
+               header_in_data = 0;
+       } else {
+               ohdr = &hdr->u.l.oth;
+               hdrsize = 8 + 40 + 12;  /* LRH + GRH + BTH */
+               /*
+                * The header with GRH is 60 bytes and the core driver sets
+                * the eager header buffer size to 56 bytes so the last 4
+                * bytes of the BTH header (PSN) is in the data buffer.
+                */
+               header_in_data =
+                       ipath_layer_get_rcvhdrentsize(dev->dd) == 16;
+               if (header_in_data) {
+                       psn = be32_to_cpu(((__be32 *) data)[0]);
+                       data += sizeof(__be32);
+               } else
+                       psn = be32_to_cpu(ohdr->bth[2]);
+       }
+       /*
+        * 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;
+
+       /*
+        * Process responses (ACKs) before anything else.  Note that the
+        * packet sequence number will be for something in the send work
+        * queue rather than the expected receive packet sequence number.
+        * In other words, this QP is the requester.
+        */
+       if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
+           opcode <= OP(ATOMIC_ACKNOWLEDGE)) {
+               ipath_rc_rcv_resp(dev, ohdr, data, tlen, qp, opcode, psn,
+                                 hdrsize, pmtu, header_in_data);
+               goto bail;
+       }
+
+       spin_lock_irqsave(&qp->r_rq.lock, flags);
+
+       /* Compute 24 bits worth of difference. */
+       diff = ipath_cmp24(psn, qp->r_psn);
+       if (unlikely(diff)) {
+               if (ipath_rc_rcv_error(dev, ohdr, data, qp, opcode,
+                                      psn, diff, header_in_data))
+                       goto done;
+               goto resched;
+       }
+
+       /* Check for opcode sequence errors. */
+       switch (qp->r_state) {
+       case OP(SEND_FIRST):
+       case OP(SEND_MIDDLE):
+               if (opcode == OP(SEND_MIDDLE) ||
+                   opcode == OP(SEND_LAST) ||
+                   opcode == OP(SEND_LAST_WITH_IMMEDIATE))
+                       break;
+       nack_inv:
+       /*
+        * A NAK will ACK earlier sends and RDMA writes.  Don't queue the
+        * NAK if a RDMA read, atomic, or NAK is pending though.
+        */
+       spin_lock(&qp->s_lock);
+       if (qp->s_ack_state >= OP(RDMA_READ_REQUEST) &&
+           qp->s_ack_state != IB_OPCODE_ACKNOWLEDGE) {
+               spin_unlock(&qp->s_lock);
+               goto done;
+       }
+       /* XXX Flush WQEs */
+       qp->state = IB_QPS_ERR;
+       qp->s_ack_state = OP(SEND_ONLY);
+       qp->s_nak_state = IB_NAK_INVALID_REQUEST;
+       qp->s_ack_psn = qp->r_psn;
+       goto resched;
+
+       case OP(RDMA_WRITE_FIRST):
+       case OP(RDMA_WRITE_MIDDLE):
+               if (opcode == OP(RDMA_WRITE_MIDDLE) ||
+                   opcode == OP(RDMA_WRITE_LAST) ||
+                   opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
+                       break;
+               goto nack_inv;
+
+       case OP(RDMA_READ_REQUEST):
+       case OP(COMPARE_SWAP):
+       case OP(FETCH_ADD):
+               /*
+                * Drop all new requests until a response has been sent.  A
+                * new request then ACKs the RDMA response we sent.  Relaxed
+                * ordering would allow new requests to be processed but we
+                * would need to keep a queue of rwqe's for all that are in
+                * progress.  Note that we can't RNR NAK this request since
+                * the RDMA READ or atomic response is already queued to be
+                * sent (unless we implement a response send queue).
+                */
+               goto done;
+
+       default:
+               if (opcode == OP(SEND_MIDDLE) ||
+                   opcode == OP(SEND_LAST) ||
+                   opcode == OP(SEND_LAST_WITH_IMMEDIATE) ||
+                   opcode == OP(RDMA_WRITE_MIDDLE) ||
+                   opcode == OP(RDMA_WRITE_LAST) ||
+                   opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
+                       goto nack_inv;
+               break;
+       }
+
+       wc.imm_data = 0;
+       wc.wc_flags = 0;
+
+       /* OK, process the packet. */
+       switch (opcode) {
+       case OP(SEND_FIRST):
+               if (!ipath_get_rwqe(qp, 0)) {
+               rnr_nak:
+                       /*
+                        * A RNR NAK will ACK earlier sends and RDMA writes.
+                        * Don't queue the NAK if a RDMA read or atomic
+                        * is pending though.
+                        */
+                       spin_lock(&qp->s_lock);
+                       if (qp->s_ack_state >=
+                           OP(RDMA_READ_REQUEST) &&
+                           qp->s_ack_state != IB_OPCODE_ACKNOWLEDGE) {
+                               spin_unlock(&qp->s_lock);
+                               goto done;
+                       }
+                       qp->s_ack_state = OP(SEND_ONLY);
+                       qp->s_nak_state = IB_RNR_NAK | qp->s_min_rnr_timer;
+                       qp->s_ack_psn = qp->r_psn;
+                       goto resched;
+               }
+               qp->r_rcv_len = 0;
+               /* FALLTHROUGH */
+       case OP(SEND_MIDDLE):
+       case OP(RDMA_WRITE_MIDDLE):
+       send_middle:
+               /* Check for invalid length PMTU or posted rwqe len. */
+               if (unlikely(tlen != (hdrsize + pmtu + 4)))
+                       goto nack_inv;
+               qp->r_rcv_len += pmtu;
+               if (unlikely(qp->r_rcv_len > qp->r_len))
+                       goto nack_inv;
+               ipath_copy_sge(&qp->r_sge, data, pmtu);
+               break;
+
+       case OP(RDMA_WRITE_LAST_WITH_IMMEDIATE):
+               /* consume RWQE */
+               if (!ipath_get_rwqe(qp, 1))
+                       goto rnr_nak;
+               goto send_last_imm;
+
+       case OP(SEND_ONLY):
+       case OP(SEND_ONLY_WITH_IMMEDIATE):
+               if (!ipath_get_rwqe(qp, 0))
+                       goto rnr_nak;
+               qp->r_rcv_len = 0;
+               if (opcode == OP(SEND_ONLY))
+                       goto send_last;
+               /* FALLTHROUGH */
+       case OP(SEND_LAST_WITH_IMMEDIATE):
+       send_last_imm:
+               if (header_in_data) {
+                       wc.imm_data = *(__be32 *) data;
+                       data += sizeof(__be32);
+               } else {
+                       /* Immediate data comes after BTH */
+                       wc.imm_data = ohdr->u.imm_data;
+               }
+               hdrsize += 4;
+               wc.wc_flags = IB_WC_WITH_IMM;
+               /* FALLTHROUGH */
+       case OP(SEND_LAST):
+       case OP(RDMA_WRITE_LAST):
+       send_last:
+               /* Get the number of bytes the message was padded by. */
+               pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
+               /* Check for invalid length. */
+               /* XXX LAST len should be >= 1 */
+               if (unlikely(tlen < (hdrsize + pad + 4)))
+                       goto nack_inv;
+               /* Don't count the CRC. */
+               tlen -= (hdrsize + pad + 4);
+               wc.byte_len = tlen + qp->r_rcv_len;
+               if (unlikely(wc.byte_len > qp->r_len))
+                       goto nack_inv;
+               ipath_copy_sge(&qp->r_sge, data, tlen);
+               atomic_inc(&qp->msn);
+               if (opcode == OP(RDMA_WRITE_LAST) ||
+                   opcode == OP(RDMA_WRITE_ONLY))
+                       break;
+               wc.wr_id = qp->r_wr_id;
+               wc.status = IB_WC_SUCCESS;
+               wc.opcode = IB_WC_RECV;
+               wc.vendor_err = 0;
+               wc.qp_num = qp->ibqp.qp_num;
+               wc.src_qp = qp->remote_qpn;
+               wc.pkey_index = 0;
+               wc.slid = qp->remote_ah_attr.dlid;
+               wc.sl = qp->remote_ah_attr.sl;
+               wc.dlid_path_bits = 0;
+               wc.port_num = 0;
+               /* Signal completion event if the solicited bit is set. */
+               ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
+                              (ohdr->bth[0] &
+                               __constant_cpu_to_be32(1 << 23)) != 0);
+               break;
+
+       case OP(RDMA_WRITE_FIRST):
+       case OP(RDMA_WRITE_ONLY):
+       case OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE):
+               /* consume RWQE */
+               /* RETH comes after BTH */
+               if (!header_in_data)
+                       reth = &ohdr->u.rc.reth;
+               else {
+                       reth = (struct ib_reth *)data;
+                       data += sizeof(*reth);
+               }
+               hdrsize += sizeof(*reth);
+               qp->r_len = be32_to_cpu(reth->length);
+               qp->r_rcv_len = 0;
+               if (qp->r_len != 0) {
+                       u32 rkey = be32_to_cpu(reth->rkey);
+                       u64 vaddr = be64_to_cpu(reth->vaddr);
+                       int ok;
+
+                       /* Check rkey & NAK */
+                       ok = ipath_rkey_ok(dev, &qp->r_sge,
+                                          qp->r_len, vaddr, rkey,
+                                          IB_ACCESS_REMOTE_WRITE);
+                       if (unlikely(!ok)) {
+                       nack_acc:
+                               /*
+                                * A NAK will ACK earlier sends and RDMA
+                                * writes.  Don't queue the NAK if a RDMA
+                                * read, atomic, or NAK is pending though.
+                                */
+                               spin_lock(&qp->s_lock);
+                               if (qp->s_ack_state >=
+                                   OP(RDMA_READ_REQUEST) &&
+                                   qp->s_ack_state !=
+                                   IB_OPCODE_ACKNOWLEDGE) {
+                                       spin_unlock(&qp->s_lock);
+                                       goto done;
+                               }
+                               /* XXX Flush WQEs */
+                               qp->state = IB_QPS_ERR;
+                               qp->s_ack_state = OP(RDMA_WRITE_ONLY);
+                               qp->s_nak_state =
+                                       IB_NAK_REMOTE_ACCESS_ERROR;
+                               qp->s_ack_psn = qp->r_psn;
+                               goto resched;
+                       }
+               } else {
+                       qp->r_sge.sg_list = NULL;
+                       qp->r_sge.sge.mr = NULL;
+                       qp->r_sge.sge.vaddr = NULL;
+                       qp->r_sge.sge.length = 0;
+                       qp->r_sge.sge.sge_length = 0;
+               }
+               if (unlikely(!(qp->qp_access_flags &
+                              IB_ACCESS_REMOTE_WRITE)))
+                       goto nack_acc;
+               if (opcode == OP(RDMA_WRITE_FIRST))
+                       goto send_middle;
+               else if (opcode == OP(RDMA_WRITE_ONLY))
+                       goto send_last;
+               if (!ipath_get_rwqe(qp, 1))
+                       goto rnr_nak;
+               goto send_last_imm;
+
+       case OP(RDMA_READ_REQUEST):
+               /* RETH comes after BTH */
+               if (!header_in_data)
+                       reth = &ohdr->u.rc.reth;
+               else {
+                       reth = (struct ib_reth *)data;
+                       data += sizeof(*reth);
+               }
+               spin_lock(&qp->s_lock);
+               if (qp->s_ack_state != OP(ACKNOWLEDGE) &&
+                   qp->s_ack_state >= IB_OPCODE_RDMA_READ_REQUEST) {
+                       spin_unlock(&qp->s_lock);
+                       goto done;
+               }
+               qp->s_rdma_len = be32_to_cpu(reth->length);
+               if (qp->s_rdma_len != 0) {
+                       u32 rkey = be32_to_cpu(reth->rkey);
+                       u64 vaddr = be64_to_cpu(reth->vaddr);
+                       int ok;
+
+                       /* Check rkey & NAK */
+                       ok = ipath_rkey_ok(dev, &qp->s_rdma_sge,
+                                          qp->s_rdma_len, vaddr, rkey,
+                                          IB_ACCESS_REMOTE_READ);
+                       if (unlikely(!ok)) {
+                               spin_unlock(&qp->s_lock);
+                               goto nack_acc;
+                       }
+                       /*
+                        * Update the next expected PSN.  We add 1 later
+                        * below, so only add the remainder here.
+                        */
+                       if (qp->s_rdma_len > pmtu)
+                               qp->r_psn += (qp->s_rdma_len - 1) / pmtu;
+               } else {
+                       qp->s_rdma_sge.sg_list = NULL;
+                       qp->s_rdma_sge.num_sge = 0;
+                       qp->s_rdma_sge.sge.mr = NULL;
+                       qp->s_rdma_sge.sge.vaddr = NULL;
+                       qp->s_rdma_sge.sge.length = 0;
+                       qp->s_rdma_sge.sge.sge_length = 0;
+               }
+               if (unlikely(!(qp->qp_access_flags &
+                              IB_ACCESS_REMOTE_READ)))
+                       goto nack_acc;
+               /*
+                * We need to increment the MSN here instead of when we
+                * finish sending the result since a duplicate request would
+                * increment it more than once.
+                */
+               atomic_inc(&qp->msn);
+               qp->s_ack_state = opcode;
+               qp->s_nak_state = 0;
+               qp->s_ack_psn = psn;
+               qp->r_psn++;
+               qp->r_state = opcode;
+               goto rdmadone;
+
+       case OP(COMPARE_SWAP):
+       case OP(FETCH_ADD): {
+               struct ib_atomic_eth *ateth;
+               u64 vaddr;
+               u64 sdata;
+               u32 rkey;
+
+               if (!header_in_data)
+                       ateth = &ohdr->u.atomic_eth;
+               else {
+                       ateth = (struct ib_atomic_eth *)data;
+                       data += sizeof(*ateth);
+               }
+               vaddr = be64_to_cpu(ateth->vaddr);
+               if (unlikely(vaddr & (sizeof(u64) - 1)))
+                       goto nack_inv;
+               rkey = be32_to_cpu(ateth->rkey);
+               /* Check rkey & NAK */
+               if (unlikely(!ipath_rkey_ok(dev, &qp->r_sge,
+                                           sizeof(u64), vaddr, rkey,
+                                           IB_ACCESS_REMOTE_ATOMIC)))
+                       goto nack_acc;
+               if (unlikely(!(qp->qp_access_flags &
+                              IB_ACCESS_REMOTE_ATOMIC)))
+                       goto nack_acc;
+               /* Perform atomic OP and save result. */
+               sdata = be64_to_cpu(ateth->swap_data);
+               spin_lock(&dev->pending_lock);
+               qp->r_atomic_data = *(u64 *) qp->r_sge.sge.vaddr;
+               if (opcode == OP(FETCH_ADD))
+                       *(u64 *) qp->r_sge.sge.vaddr =
+                               qp->r_atomic_data + sdata;
+               else if (qp->r_atomic_data ==
+                        be64_to_cpu(ateth->compare_data))
+                       *(u64 *) qp->r_sge.sge.vaddr = sdata;
+               spin_unlock(&dev->pending_lock);
+               atomic_inc(&qp->msn);
+               qp->r_atomic_psn = psn & IPS_PSN_MASK;
+               psn |= 1 << 31;
+               break;
+       }
+
+       default:
+               /* Drop packet for unknown opcodes. */
+               goto done;
+       }
+       qp->r_psn++;
+       qp->r_state = opcode;
+       /* Send an ACK if requested or required. */
+       if (psn & (1 << 31)) {
+               /*
+                * Coalesce ACKs unless there is a RDMA READ or
+                * ATOMIC pending.
+                */
+               spin_lock(&qp->s_lock);
+               if (qp->s_ack_state == OP(ACKNOWLEDGE) ||
+                   qp->s_ack_state < IB_OPCODE_RDMA_READ_REQUEST) {
+                       qp->s_ack_state = opcode;
+                       qp->s_nak_state = 0;
+                       qp->s_ack_psn = psn;
+                       qp->s_ack_atomic = qp->r_atomic_data;
+                       goto resched;
+               }
+               spin_unlock(&qp->s_lock);
+       }
+done:
+       spin_unlock_irqrestore(&qp->r_rq.lock, flags);
+       goto bail;
+
+resched:
+       /*
+        * Try to send ACK right away but not if ipath_do_rc_send() is
+        * active.
+        */
+       if (qp->s_hdrwords == 0 &&
+           (qp->s_ack_state < IB_OPCODE_RDMA_READ_REQUEST ||
+            qp->s_ack_state >= IB_OPCODE_COMPARE_SWAP))
+               send_rc_ack(qp);
+
+rdmadone:
+       spin_unlock(&qp->s_lock);
+       spin_unlock_irqrestore(&qp->r_rq.lock, flags);
+
+       /* Call ipath_do_rc_send() in another thread. */
+       tasklet_hi_schedule(&qp->s_task);
+
+bail:
+       return;
+}