IB/ipath: infiniband RC protocol support

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>

1 files changed, 1857 insertions, 0 deletions

diff --git a/drivers/infiniband/hw/ipath/ipath_rc.c b/drivers/infiniband/hw/ipath/ipath_rc.c
new file mode 100644
index 00000000000..a4055ca0061
--- /dev/null
+++ b/drivers/infiniband/hw/ipath/ipath_rc.c
@@ -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_