IB/qib: Add new qib driver for QLogic PCIe InfiniBand adapters

Add a low-level IB driver for QLogic PCIe adapters.

Signed-off-by: Ralph Campbell <ralph.campbell@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>

1 files changed, 2288 insertions, 0 deletions

diff --git a/drivers/infiniband/hw/qib/qib_rc.c b/drivers/infiniband/hw/qib/qib_rc.c
new file mode 100644
index 00000000000..40c0a373719
--- /dev/null
+++ b/drivers/infiniband/hw/qib/qib_rc.c
@@ -0,0 +1,2288 @@
+/*
+ * Copyright (c) 2006, 2007, 2008, 2009 QLogic Corporation. All rights reserved.
+ * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include <linux/io.h>
+
+#include "qib.h"
+
+/* cut down ridiculously long IB macro names */
+#define OP(x) IB_OPCODE_RC_##x
+
+static void rc_timeout(unsigned long arg);
+
+static u32 restart_sge(struct qib_sge_state *ss, struct qib_swqe *wqe,
+		       u32 psn, u32 pmtu)
+{
+	u32 len;
+
+	len = ((psn - wqe->psn) & QIB_PSN_MASK) * pmtu;
+	ss->sge = wqe->sg_list[0];
+	ss->sg_list = wqe->sg_list + 1;
+	ss->num_sge = wqe->wr.num_sge;
+	ss->total_len = wqe->length;
+	qib_skip_sge(ss, len, 0);
+	return wqe->length - len;
+}
+
+static void start_timer(struct qib_qp *qp)
+{
+	qp->s_flags |= QIB_S_TIMER;
+	qp->s_timer.function = rc_timeout;
+	/* 4.096 usec. * (1 << qp->timeout) */
+	qp->s_timer.expires = jiffies +
+		usecs_to_jiffies((4096UL * (1UL << qp->timeout)) / 1000UL);
+	add_timer(&qp->s_timer);
+}
+
+/**
+ * qib_make_rc_ack - construct a response packet (ACK, NAK, or RDMA read)
+ * @dev: the device for this QP
+ * @qp: a pointer to the QP
+ * @ohdr: a pointer to the IB header being constructed
+ * @pmtu: the path MTU
+ *
+ * Return 1 if constructed; otherwise, return 0.
+ * Note that we are in the responder's side of the QP context.
+ * Note the QP s_lock must be held.
+ */
+static int qib_make_rc_ack(struct qib_ibdev *dev, struct qib_qp *qp,
+			   struct qib_other_headers *ohdr, u32 pmtu)
+{
+	struct qib_ack_entry *e;
+	u32 hwords;
+	u32 len;
+	u32 bth0;
+	u32 bth2;
+
+	/* Don't send an ACK if we aren't supposed to. */
+	if (!(ib_qib_state_ops[qp->state] & QIB_PROCESS_RECV_OK))
+		goto bail;
+
+	/* header size in 32-bit words LRH+BTH = (8+12)/4. */
+	hwords = 5;
+
+	switch (qp->s_ack_state) {
+	case OP(RDMA_READ_RESPONSE_LAST):
+	case OP(RDMA_READ_RESPONSE_ONLY):
+		e = &qp->s_ack_queue[qp->s_tail_ack_queue];
+		if (e->rdma_sge.mr) {
+			atomic_dec(&e->rdma_sge.mr->refcount);
+			e->rdma_sge.mr = NULL;
+		}
+		/* FALLTHROUGH */
+	case OP(ATOMIC_ACKNOWLEDGE):
+		/*
+		 * We can increment the tail pointer now that the last
+		 * response has been sent instead of only being
+		 * constructed.
+		 */
+		if (++qp->s_tail_ack_queue > QIB_MAX_RDMA_ATOMIC)
+			qp->s_tail_ack_queue = 0;
+		/* FALLTHROUGH */
+	case OP(SEND_ONLY):
+	case OP(ACKNOWLEDGE):
+		/* Check for no next entry in the queue. */
+		if (qp->r_head_ack_queue == qp->s_tail_ack_queue) {
+			if (qp->s_flags & QIB_S_ACK_PENDING)
+				goto normal;
+			goto bail;
+		}
+
+		e = &qp->s_ack_queue[qp->s_tail_ack_queue];
+		if (e->opcode == OP(RDMA_READ_REQUEST)) {
+			/*
+			 * If a RDMA read response is being resent and
+			 * we haven't seen the duplicate request yet,
+			 * then stop sending the remaining responses the
+			 * responder has seen until the requester resends it.
+			 */
+			len = e->rdma_sge.sge_length;
+			if (len && !e->rdma_sge.mr) {
+				qp->s_tail_ack_queue = qp->r_head_ack_queue;
+				goto bail;
+			}
+			/* Copy SGE state in case we need to resend */
+			qp->s_rdma_mr = e->rdma_sge.mr;
+			if (qp->s_rdma_mr)
+				atomic_inc(&qp->s_rdma_mr->refcount);
+			qp->s_ack_rdma_sge.sge = e->rdma_sge;
+			qp->s_ack_rdma_sge.num_sge = 1;
+			qp->s_cur_sge = &qp->s_ack_rdma_sge;
+			if (len > pmtu) {
+				len = pmtu;
+				qp->s_ack_state = OP(RDMA_READ_RESPONSE_FIRST);
+			} else {
+				qp->s_ack_state = OP(RDMA_READ_RESPONSE_ONLY);
+				e->sent = 1;
+			}
+			ohdr->u.aeth = qib_compute_aeth(qp);
+			hwords++;
+			qp->s_ack_rdma_psn = e->psn;
+			bth2 = qp->s_ack_rdma_psn++ & QIB_PSN_MASK;
+		} else {
+			/* COMPARE_SWAP or FETCH_ADD */
+			qp->s_cur_sge = NULL;
+			len = 0;
+			qp->s_ack_state = OP(ATOMIC_ACKNOWLEDGE);
+			ohdr->u.at.aeth = qib_compute_aeth(qp);
+			ohdr->u.at.atomic_ack_eth[0] =
+				cpu_to_be32(e->atomic_data >> 32);
+			ohdr->u.at.atomic_ack_eth[1] =
+				cpu_to_be32(e->atomic_data);
+			hwords += sizeof(ohdr->u.at) / sizeof(u32);
+			bth2 = e->psn & QIB_PSN_MASK;
+			e->sent = 1;
+		}
+		bth0 = qp->s_ack_state << 24;
+		break;
+
+	case OP(RDMA_READ_RESPONSE_FIRST):
+		qp->s_ack_state = OP(RDMA_READ_RESPONSE_MIDDLE);
+		/* FALLTHROUGH */
+	case OP(RDMA_READ_RESPONSE_MIDDLE):
+		qp->s_cur_sge = &qp->s_ack_rdma_sge;
+		qp->s_rdma_mr = qp->s_ack_rdma_sge.sge.mr;
+		if (qp->s_rdma_mr)
+			atomic_inc(&qp->s_rdma_mr->refcount);
+		len = qp->s_ack_rdma_sge.sge.sge_length;
+		if (len > pmtu)
+			len = pmtu;
+		else {
+			ohdr->u.aeth = qib_compute_aeth(qp);
+			hwords++;
+			qp->s_ack_state = OP(RDMA_READ_RESPONSE_LAST);
+			e = &qp->s_ack_queue[qp->s_tail_ack_queue];
+			e->sent = 1;
+		}
+		bth0 = qp->s_ack_state << 24;
+		bth2 = qp->s_ack_rdma_psn++ & QIB_PSN_MASK;
+		break;
+
+	default:
+normal:
+		/*
+		 * Send a regular ACK.
+		 * Set the s_ack_state so we wait until after sending
+		 * the ACK before setting s_ack_state to ACKNOWLEDGE
+		 * (see above).
+		 */
+		qp->s_ack_state = OP(SEND_ONLY);
+		qp->s_flags &= ~QIB_S_ACK_PENDING;
+		qp->s_cur_sge = NULL;
+		if (qp->s_nak_state)
+			ohdr->u.aeth =
+				cpu_to_be32((qp->r_msn & QIB_MSN_MASK) |
+					    (qp->s_nak_state <<
+					     QIB_AETH_CREDIT_SHIFT));
+		else
+			ohdr->u.aeth = qib_compute_aeth(qp);
+		hwords++;
+		len = 0;
+		bth0 = OP(ACKNOWLEDGE) << 24;
+		bth2 = qp->s_ack_psn & QIB_PSN_MASK;
+	}
+	qp->s_rdma_ack_cnt++;
+	qp->s_hdrwords = hwords;
+	qp->s_cur_size = len;
+	qib_make_ruc_header(qp, ohdr, bth0, bth2);
+	return 1;
+
+bail:
+	qp->s_ack_state = OP(ACKNOWLEDGE);
+	qp->s_flags &= ~(QIB_S_RESP_PENDING | QIB_S_ACK_PENDING);
+	return 0;
+}
+
+/**
+ * qib_make_rc_req - construct a request packet (SEND, RDMA r/w, ATOMIC)
+ * @qp: a pointer to the QP
+ *
+ * Return 1 if constructed; otherwise, return 0.
+ */
+int qib_make_rc_req(struct qib_qp *qp)
+{
+	struct qib_ibdev *dev = to_idev(qp->ibqp.device);
+	struct qib_other_headers *ohdr;
+	struct qib_sge_state *ss;
+	struct qib_swqe *wqe;
+	u32 hwords;
+	u32 len;
+	u32 bth0;
+	u32 bth2;
+	u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
+	char newreq;
+	unsigned long flags;
+	int ret = 0;
+	int delta;
+
+	ohdr = &qp->s_hdr.u.oth;
+	if (qp->remote_ah_attr.ah_flags & IB_AH_GRH)
+		ohdr = &qp->s_hdr.u.l.oth;
+
+	/*
+	 * The lock is needed to synchronize between the sending tasklet,
+	 * the receive interrupt handler, and timeout resends.
+	 */
+	spin_lock_irqsave(&qp->s_lock, flags);
+
+	/* Sending responses has higher priority over sending requests. */
+	if ((qp->s_flags & QIB_S_RESP_PENDING) &&
+	    qib_make_rc_ack(dev, qp, ohdr, pmtu))
+		goto done;
+
+	if (!(ib_qib_state_ops[qp->state] & QIB_PROCESS_SEND_OK)) {
+		if (!(ib_qib_state_ops[qp->state] & QIB_FLUSH_SEND))
+			goto bail;
+		/* We are in the error state, flush the work request. */
+		if (qp->s_last == qp->s_head)
+			goto bail;
+		/* If DMAs are in progress, we can't flush immediately. */
+		if (atomic_read(&qp->s_dma_busy)) {
+			qp->s_flags |= QIB_S_WAIT_DMA;
+			goto bail;
+		}
+		wqe = get_swqe_ptr(qp, qp->s_last);
+		while (qp->s_last != qp->s_acked) {
+			qib_send_complete(qp, wqe, IB_WC_SUCCESS);
+			if (++qp->s_last >= qp->s_size)
+				qp->s_last = 0;
+			wqe = get_swqe_ptr(qp, qp->s_last);
+		}
+		qib_send_complete(qp, wqe, IB_WC_WR_FLUSH_ERR);
+		goto done;
+	}
+
+	if (qp->s_flags & (QIB_S_WAIT_RNR | QIB_S_WAIT_ACK))
+		goto bail;
+
+	if (qib_cmp24(qp->s_psn, qp->s_sending_hpsn) <= 0) {
+		if (qib_cmp24(qp->s_sending_psn, qp->s_sending_hpsn) <= 0) {
+			qp->s_flags |= QIB_S_WAIT_PSN;
+			goto bail;
+		}
+		qp->s_sending_psn = qp->s_psn;
+		qp->s_sending_hpsn = qp->s_psn - 1;
+	}
+
+	/* 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:
+		if (!(ib_qib_state_ops[qp->state] & QIB_PROCESS_NEXT_SEND_OK))
+			goto bail;
+		/*
+		 * 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 bail;
+			/*
+			 * If a fence is requested, wait for previous
+			 * RDMA read and atomic operations to finish.
+			 */
+			if ((wqe->wr.send_flags & IB_SEND_FENCE) &&
+			    qp->s_num_rd_atomic) {
+				qp->s_flags |= QIB_S_WAIT_FENCE;
+				goto bail;
+			}
+			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.
+		 */
+		len = wqe->length;
+		ss = &qp->s_sge;
+		bth2 = qp->s_psn & QIB_PSN_MASK;
+		switch (wqe->wr.opcode) {
+		case IB_WR_SEND:
+		case IB_WR_SEND_WITH_IMM:
+			/* If no credit, return. */
+			if (!(qp->s_flags & QIB_S_UNLIMITED_CREDIT) &&
+			    qib_cmp24(wqe->ssn, qp->s_lsn + 1) > 0) {
+				qp->s_flags |= QIB_S_WAIT_SSN_CREDIT;
+				goto bail;
+			}
+			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.ex.imm_data;
+				hwords += 1;
+			}
+			if (wqe->wr.send_flags & IB_SEND_SOLICITED)
+				bth0 |= IB_BTH_SOLICITED;
+			bth2 |= IB_BTH_REQ_ACK;
+			if (++qp->s_cur == qp->s_size)
+				qp->s_cur = 0;
+			break;
+
+		case IB_WR_RDMA_WRITE:
+			if (newreq && !(qp->s_flags & QIB_S_UNLIMITED_CREDIT))
+				qp->s_lsn++;
+			/* FALLTHROUGH */
+		case IB_WR_RDMA_WRITE_WITH_IMM:
+			/* If no credit, return. */
+			if (!(qp->s_flags & QIB_S_UNLIMITED_CREDIT) &&
+			    qib_cmp24(wqe->ssn, qp->s_lsn + 1) > 0) {
+				qp->s_flags |= QIB_S_WAIT_SSN_CREDIT;
+				goto bail;
+			}
+			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) / sizeof(u32);
+			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.ex.imm_data;
+				hwords += 1;
+				if (wqe->wr.send_flags & IB_SEND_SOLICITED)
+					bth0 |= IB_BTH_SOLICITED;
+			}
+			bth2 |= IB_BTH_REQ_ACK;
+			if (++qp->s_cur == qp->s_size)
+				qp->s_cur = 0;
+			break;
+
+		case IB_WR_RDMA_READ:
+			/*
+			 * Don't allow more operations to be started
+			 * than the QP limits allow.
+			 */
+			if (newreq) {
+				if (qp->s_num_rd_atomic >=
+				    qp->s_max_rd_atomic) {
+					qp->s_flags |= QIB_S_WAIT_RDMAR;
+					goto bail;
+				}
+				qp->s_num_rd_atomic++;
+				if (!(qp->s_flags & QIB_S_UNLIMITED_CREDIT))
+					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++;
+			}
+			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) / sizeof(u32);
+			ss = NULL;
+			len = 0;
+			bth2 |= IB_BTH_REQ_ACK;
+			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:
+			/*
+			 * Don't allow more operations to be started
+			 * than the QP limits allow.
+			 */
+			if (newreq) {
+				if (qp->s_num_rd_atomic >=
+				    qp->s_max_rd_atomic) {
+					qp->s_flags |= QIB_S_WAIT_RDMAR;
+					goto bail;
+				}
+				qp->s_num_rd_atomic++;
+				if (!(qp->s_flags & QIB_S_UNLIMITED_CREDIT))
+					qp->s_lsn++;
+				wqe->lpsn = wqe->psn;
+			}
+			if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
+				qp->s_state = OP(COMPARE_SWAP);
+				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);
+			} else {
+				qp->s_state = OP(FETCH_ADD);
+				ohdr->u.atomic_eth.swap_data = cpu_to_be64(
+					wqe->wr.wr.atomic.compare_add);
+				ohdr->u.atomic_eth.compare_data = 0;
+			}
+			ohdr->u.atomic_eth.vaddr[0] = cpu_to_be32(
+				wqe->wr.wr.atomic.remote_addr >> 32);
+			ohdr->u.atomic_eth.vaddr[1] = cpu_to_be32(
+				wqe->wr.wr.atomic.remote_addr);
+			ohdr->u.atomic_eth.rkey = cpu_to_be32(
+				wqe->wr.wr.atomic.rkey);
+			hwords += sizeof(struct ib_atomic_eth) / sizeof(u32);
+			ss = NULL;
+			len = 0;
+			bth2 |= IB_BTH_REQ_ACK;
+			if (++qp->s_cur == qp->s_size)
+				qp->s_cur = 0;
+			break;
+
+		default:
+			goto bail;
+		}
+		qp->s_sge.sge = wqe->sg_list[0];
+		qp->s_sge.sg_list = wqe->sg_list + 1;
+		qp->s_sge.num_sge = wqe->wr.num_sge;
+		qp->s_sge.total_len = wqe->length;
+		qp->s_len = wqe->length;
+		if (newreq) {
+			qp->s_tail++;
+			if (qp->s_tail >= qp->s_size)
+				qp->s_tail = 0;
+		}
+		if (wqe->wr.opcode == IB_WR_RDMA_READ)
+			qp->s_psn = wqe->lpsn + 1;
+		else {
+			qp->s_psn++;
+			if (qib_cmp24(qp->s_psn, qp->s_next_psn) > 0)
+				qp->s_next_psn = qp->s_psn;
+		}
+		break;
+
+	case OP(RDMA_READ_RESPONSE_FIRST):
+		/*
+		 * qp->s_state is normally set to the opcode of the
+		 * last packet constructed for new requests and therefore
+		 * is never set to RDMA read response.
+		 * RDMA_READ_RESPONSE_FIRST is used by the ACK processing
+		 * thread to indicate a SEND needs to be restarted from an
+		 * earlier PSN without interferring with the sending thread.
+		 * See qib_restart_rc().
+		 */
+		qp->s_len = restart_sge(&qp->s_sge, wqe, qp->s_psn, pmtu);
+		/* FALLTHROUGH */
+	case OP(SEND_FIRST):
+		qp->s_state = OP(SEND_MIDDLE);
+		/* FALLTHROUGH */
+	case OP(SEND_MIDDLE):
+		bth2 = qp->s_psn++ & QIB_PSN_MASK;
+		if (qib_cmp24(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) {
+			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.ex.imm_data;
+			hwords += 1;
+		}
+		if (wqe->wr.send_flags & IB_SEND_SOLICITED)
+			bth0 |= IB_BTH_SOLICITED;
+		bth2 |= IB_BTH_REQ_ACK;
+		qp->s_cur++;
+		if (qp->s_cur >= qp->s_size)
+			qp->s_cur = 0;
+		break;
+
+	case OP(RDMA_READ_RESPONSE_LAST):
+		/*
+		 * qp->s_state is normally set to the opcode of the
+		 * last packet constructed for new requests and therefore
+		 * is never set to RDMA read response.
+		 * RDMA_READ_RESPONSE_LAST is used by the ACK processing
+		 * thread to indicate a RDMA write needs to be restarted from
+		 * an earlier PSN without interferring with the sending thread.
+		 * See qib_restart_rc().
+		 */
+		qp->s_len = restart_sge(&qp->s_sge, wqe, qp->s_psn, pmtu);
+		/* FALLTHROUGH */
+	case OP(RDMA_WRITE_FIRST):
+		qp->s_state = OP(RDMA_WRITE_MIDDLE);
+		/* FALLTHROUGH */
+	case OP(RDMA_WRITE_MIDDLE):
+		bth2 = qp->s_psn++ & QIB_PSN_MASK;
+		if (qib_cmp24(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) {
+			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.ex.imm_data;
+			hwords += 1;
+			if (wqe->wr.send_flags & IB_SEND_SOLICITED)
+				bth0 |= IB_BTH_SOLICITED;
+		}
+		bth2 |= IB_BTH_REQ_ACK;
+		qp->s_cur++;
+		if (qp->s_cur >= qp->s_size)
+			qp->s_cur = 0;
+		break;
+
+	case OP(RDMA_READ_RESPONSE_MIDDLE):
+		/*
+		 * qp->s_state is normally set to the opcode of the
+		 * last packet constructed for new requests and therefore
+		 * is never set to RDMA read response.
+		 * RDMA_READ_RESPONSE_MIDDLE is used by the ACK processing
+		 * thread to indicate a RDMA read needs to be restarted from
+		 * an earlier PSN without interferring with the sending thread.
+		 * See qib_restart_rc().
+		 */
+		len = ((qp->s_psn - wqe->psn) & QIB_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(wqe->length - len);
+		qp->s_state = OP(RDMA_READ_REQUEST);
+		hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
+		bth2 = (qp->s_psn & QIB_PSN_MASK) | IB_BTH_REQ_ACK;
+		qp->s_psn = wqe->lpsn + 1;
+		ss = NULL;
+		len = 0;
+		qp->s_cur++;
+		if (qp->s_cur == qp->s_size)
+			qp->s_cur = 0;
+		break;
+	}
+	qp->s_sending_hpsn = bth2;
+	delta = (((int) bth2 - (int) wqe->psn) << 8) >> 8;
+	if (delta && delta % QIB_PSN_CREDIT == 0)
+		bth2 |= IB_BTH_REQ_ACK;
+	if (qp->s_flags & QIB_S_SEND_ONE) {
+		qp->s_flags &= ~QIB_S_SEND_ONE;
+		qp->s_flags |= QIB_S_WAIT_ACK;
+		bth2 |= IB_BTH_REQ_ACK;
+	}
+	qp->s_len -= len;
+	qp->s_hdrwords = hwords;
+	qp->s_cur_sge = ss;
+	qp->s_cur_size = len;
+	qib_make_ruc_header(qp, ohdr, bth0 | (qp->s_state << 24), bth2);
+done:
+	ret = 1;
+	goto unlock;
+
+bail:
+	qp->s_flags &= ~QIB_S_BUSY;
+unlock:
+	spin_unlock_irqrestore(&qp->s_lock, flags);
+	return ret;
+}
+
+/**
+ * qib_send_rc_ack - Construct an ACK packet and send it
+ * @qp: a pointer to the QP
+ *
+ * This is called from qib_rc_rcv() and qib_kreceive().
+ * Note that RDMA reads and atomics are handled in the
+ * send side QP state and tasklet.
+ */
+void qib_send_rc_ack(struct qib_qp *qp)
+{
+	struct qib_devdata *dd = dd_from_ibdev(qp->ibqp.device);
+	struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
+	struct qib_pportdata *ppd = ppd_from_ibp(ibp);
+	u64 pbc;
+	u16 lrh0;
+	u32 bth0;
+	u32 hwords;
+	u32 pbufn;
+	u32 __iomem *piobuf;
+	struct qib_ib_header hdr;
+	struct qib_other_headers *ohdr;
+	u32 control;
+	unsigned long flags;
+
+	spin_lock_irqsave(&qp->s_lock, flags);
+
+	if (!(ib_qib_state_ops[qp->state] & QIB_PROCESS_RECV_OK))
+		goto unlock;
+
+	/* Don't send ACK or NAK if a RDMA read or atomic is pending. */
+	if ((qp->s_flags & QIB_S_RESP_PENDING) || qp->s_rdma_ack_cnt)
+		goto queue_ack;
+
+	/* Construct the header with s_lock held so APM doesn't change it. */
+	ohdr = &hdr.u.oth;
+	lrh0 = QIB_LRH_BTH;
+	/* header size in 32-bit words LRH+BTH+AETH = (8+12+4)/4. */
+	hwords = 6;
+	if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
+		hwords += qib_make_grh(ibp, &hdr.u.l.grh,
+				       &qp->remote_ah_attr.grh, hwords, 0);
+		ohdr = &hdr.u.l.oth;
+		lrh0 = QIB_LRH_GRH;
+	}
+	/* read pkey_index w/o lock (its atomic) */
+	bth0 = qib_get_pkey(ibp, qp->s_pkey_index) | (OP(ACKNOWLEDGE) << 24);
+	if (qp->s_mig_state == IB_MIG_MIGRATED)
+		bth0 |= IB_BTH_MIG_REQ;
+	if (qp->r_nak_state)
+		ohdr->u.aeth = cpu_to_be32((qp->r_msn & QIB_MSN_MASK) |
+					    (qp->r_nak_state <<
+					     QIB_AETH_CREDIT_SHIFT));
+	else
+		ohdr->u.aeth = qib_compute_aeth(qp);
+	lrh0 |= ibp->sl_to_vl[qp->remote_ah_attr.sl] << 12 |
+		qp->remote_ah_attr.sl << 4;
+	hdr.lrh[0] = cpu_to_be16(lrh0);
+	hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
+	hdr.lrh[2] = cpu_to_be16(hwords + SIZE_OF_CRC);
+	hdr.lrh[3] = cpu_to_be16(ppd->lid | qp->remote_ah_attr.src_path_bits);
+	ohdr->bth[0] = cpu_to_be32(bth0);
+	ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
+	ohdr->bth[2] = cpu_to_be32(qp->r_ack_psn & QIB_PSN_MASK);
+
+	spin_unlock_irqrestore(&qp->s_lock, flags);
+
+	/* Don't try to send ACKs if the link isn't ACTIVE */
+	if (!(ppd->lflags & QIBL_LINKACTIVE))
+		goto done;
+
+	control = dd->f_setpbc_control(ppd, hwords + SIZE_OF_CRC,
+				       qp->s_srate, lrh0 >> 12);
+	/* length is + 1 for the control dword */
+	pbc = ((u64) control << 32) | (hwords + 1);
+
+	piobuf = dd->f_getsendbuf(ppd, pbc, &pbufn);
+	if (!piobuf) {
+		/*
+		 * We are out of PIO buffers at the moment.
+		 * Pass responsibility for sending the ACK to the
+		 * send tasklet so that when a PIO buffer becomes
+		 * available, the ACK is sent ahead of other outgoing
+		 * packets.
+		 */
+		spin_lock_irqsave(&qp->s_lock, flags);
+		goto queue_ack;
+	}
+
+	/*
+	 * Write the pbc.
+	 * We have to flush after the PBC for correctness
+	 * on some cpus or WC buffer can be written out of order.
+	 */
+	writeq(pbc, piobuf);
+
+	if (dd->flags & QIB_PIO_FLUSH_WC) {
+		u32 *hdrp = (u32 *) &hdr;
+
+		qib_flush_wc();
+		qib_pio_copy(piobuf + 2, hdrp, hwords - 1);
+		qib_flush_wc();
+		__raw_writel(hdrp[hwords - 1], piobuf + hwords + 1);
+	} else
+		qib_pio_copy(piobuf + 2, (u32 *) &hdr, hwords);
+
+	if (dd->flags & QIB_USE_SPCL_TRIG) {
+		u32 spcl_off = (pbufn >= dd->piobcnt2k) ? 2047 : 1023;
+
+		qib_flush_wc();
+		__raw_writel(0xaebecede, piobuf + spcl_off);
+	}
+
+	qib_flush_wc();
+	qib_sendbuf_done(dd, pbufn);
+
+	ibp->n_unicast_xmit++;
+	goto done;
+
+queue_ack:
+	if (ib_qib_state_ops[qp->state] & QIB_PROCESS_RECV_OK) {
+		ibp->n_rc_qacks++;
+		qp->s_flags |= QIB_S_ACK_PENDING | QIB_S_RESP_PENDING;
+		qp->s_nak_state = qp->r_nak_state;
+		qp->s_ack_psn = qp->r_ack_psn;
+
+		/* Schedule the send tasklet. */
+		qib_schedule_send(qp);
+	}
+unlock:
+	spin_unlock_irqrestore(&qp->s_lock, flags);
+done:
+	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 qib_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 qib_qp *qp, u32 psn)
+{
+	u32 n = qp->s_acked;
+	struct qib_swqe *wqe = get_swqe_ptr(qp, n);
+	u32 opcode;
+
+	qp->s_cur = n;
+
+	/*
+	 * If we are starting the request from the beginning,
+	 * let the normal send code handle initialization.
+	 */
+	if (qib_cmp24(psn, wqe->psn) <= 0) {
+		qp->s_state = OP(SEND_LAST);
+		goto done;
+	}
+
+	/* Find the work request opcode corresponding to the given PSN. */
+	opcode = wqe->wr.opcode;
+	for (;;) {
+		int diff;
+
+		if (++n == qp->s_size)
+			n = 0;
+		if (n == qp->s_tail)
+			break;
+		wqe = get_swqe_ptr(qp, n);
+		diff = qib_cmp24(psn, wqe->psn);
+		if (diff < 0)
+			break;
+		qp->s_cur = n;
+		/*
+		 * If we are starting the request from the beginning,
+		 * let the normal send code handle initialization.
+		 */
+		if (diff == 0) {
+			qp->s_state = OP(SEND_LAST);
+			goto done;
+		}
+		opcode = wqe->wr.opcode;
+	}
+
+	/*
+	 * 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 qib_make_rc_req().
+	 */
+	switch (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:
+	qp->s_psn = psn;
+	/*
+	 * Set QIB_S_WAIT_PSN as qib_rc_complete() may start the timer
+	 * asynchronously before the send tasklet can get scheduled.
+	 * Doing it in qib_make_rc_req() is too late.
+	 */
+	if ((qib_cmp24(qp->s_psn, qp->s_sending_hpsn) <= 0) &&
+	    (qib_cmp24(qp->s_sending_psn, qp->s_sending_hpsn) <= 0))
+		qp->s_flags |= QIB_S_WAIT_PSN;
+}
+
+/*
+ * Back up requester to resend the last un-ACKed request.
+ * The QP s_lock should be held and interrupts disabled.
+ */
+static void qib_restart_rc(struct qib_qp *qp, u32 psn, int wait)
+{
+	struct qib_swqe *wqe = get_swqe_ptr(qp, qp->s_acked);
+	struct qib_ibport *ibp;
+
+	if (qp->s_retry == 0) {
+		if (qp->s_mig_state == IB_MIG_ARMED) {
+			qib_migrate_qp(qp);
+			qp->s_retry = qp->s_retry_cnt;
+		} else if (qp->s_last == qp->s_acked) {
+			qib_send_complete(qp, wqe, IB_WC_RETRY_EXC_ERR);
+			qib_error_qp(qp, IB_WC_WR_FLUSH_ERR);
+			return;
+		} else /* XXX need to handle delayed completion */
+			return;
+	} else
+		qp->s_retry--;
+
+	ibp = to_iport(qp->ibqp.device, qp->port_num);
+	if (wqe->wr.opcode == IB_WR_RDMA_READ)
+		ibp->n_rc_resends++;
+	else
+		ibp->n_rc_resends += (qp->s_psn - psn) & QIB_PSN_MASK;
+
+	qp->s_flags &= ~(QIB_S_WAIT_FENCE | QIB_S_WAIT_RDMAR |
+			 QIB_S_WAIT_SSN_CREDIT | QIB_S_WAIT_PSN |
+			 QIB_S_WAIT_ACK);
+	if (wait)
+		qp->s_flags |= QIB_S_SEND_ONE;
+	reset_psn(qp, psn);
+}
+
+/*
+ * This is called from s_timer for missing responses.
+ */
+static void rc_timeout(unsigned long arg)
+{
+	struct qib_qp *qp = (struct qib_qp *)arg;
+	struct qib_ibport *ibp;
+	unsigned long flags;
+
+	spin_lock_irqsave(&qp->s_lock, flags);
+	if (qp->s_flags & QIB_S_TIMER) {
+		ibp = to_iport(qp->ibqp.device, qp->port_num);
+		ibp->n_rc_timeouts++;
+		qp->s_flags &= ~QIB_S_TIMER;
+		del_timer(&qp->s_timer);
+		qib_restart_rc(qp, qp->s_last_psn + 1, 1);
+		qib_schedule_send(qp);
+	}
+	spin_unlock_irqrestore(&qp->s_lock, flags);
+}
+
+/*
+ * This is called from s_timer for RNR timeouts.
+ */
+void qib_rc_rnr_retry(unsigned long arg)
+{
+	struct qib_qp *qp = (struct qib_qp *)arg;
+	unsigned long flags;
+
+	spin_lock_irqsave(&qp->s_lock, flags);
+	if (qp->s_flags & QIB_S_WAIT_RNR) {
+		qp->s_flags &= ~QIB_S_WAIT_RNR;
+		del_timer(&qp->s_timer);
+		qib_schedule_send(qp);
+	}
+	spin_unlock_irqrestore(&qp->s_lock, flags);
+}
+
+/*
+ * Set qp->s_sending_psn to the next PSN after the given one.
+ * This would be psn+1 except when RDMA reads are present.
+ */
+static void reset_sending_psn(struct qib_qp *qp, u32 psn)
+{
+	struct qib_swqe *wqe;
+	u32 n = qp->s_last;
+
+	/* Find the work request corresponding to the given PSN. */
+	for (;;) {
+		wqe = get_swqe_ptr(qp, n);
+		if (qib_cmp24(psn, wqe->lpsn) <= 0) {
+			if (wqe->wr.opcode == IB_WR_RDMA_READ)
+				qp->s_sending_psn = wqe->lpsn + 1;
+			else
+				qp->s_sending_psn = psn + 1;
+			break;
+		}
+		if (++n == qp->s_size)
+			n = 0;
+		if (n == qp->s_tail)
+			break;
+	}
+}
+
+/*
+ * This should be called with the QP s_lock held and interrupts disabled.
+ */
+void qib_rc_send_complete(struct qib_qp *qp, struct qib_ib_header *hdr)
+{
+	struct qib_other_headers *ohdr;
+	struct qib_swqe *wqe;
+	struct ib_wc wc;
+	unsigned i;
+	u32 opcode;
+	u32 psn;
+
+	if (!(ib_qib_state_ops[qp->state] & QIB_PROCESS_OR_FLUSH_SEND))
+		return;
+
+	/* Find out where the BTH is */
+	if ((be16_to_cpu(hdr->lrh[0]) & 3) == QIB_LRH_BTH)
+		ohdr = &hdr->u.oth;
+	else
+		ohdr = &hdr->u.l.oth;
+
+	opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
+	if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
+	    opcode <= OP(ATOMIC_ACKNOWLEDGE)) {
+		WARN_ON(!qp->s_rdma_ack_cnt);
+		qp->s_rdma_ack_cnt--;
+		return;
+	}
+
+	psn = be32_to_cpu(ohdr->bth[2]);
+	reset_sending_psn(qp, psn);
+
+	/*
+	 * Start timer after a packet requesting an ACK has been sent and
+	 * there are still requests that haven't been acked.
+	 */
+	if ((psn & IB_BTH_REQ_ACK) && qp->s_acked != qp->s_tail &&
+	    !(qp->s_flags & (QIB_S_TIMER | QIB_S_WAIT_RNR | QIB_S_WAIT_PSN)))
+		start_timer(qp);
+
+	while (qp->s_last != qp->s_acked) {
+		wqe = get_swqe_ptr(qp, qp->s_last);
+		if (qib_cmp24(wqe->lpsn, qp->s_sending_psn) >= 0 &&
+		    qib_cmp24(qp->s_sending_psn, qp->s_sending_hpsn) <= 0)
+			break;
+		for (i = 0; i < wqe->wr.num_sge; i++) {
+			struct qib_sge *sge = &wqe->sg_list[i];
+
+			atomic_dec(&sge->mr->refcount);
+		}
+		/* Post a send completion queue entry if requested. */
+		if (!(qp->s_flags & QIB_S_SIGNAL_REQ_WR) ||
+		    (wqe->wr.send_flags & IB_SEND_SIGNALED)) {
+			memset(&wc, 0, sizeof wc);
+			wc.wr_id = wqe->wr.wr_id;
+			wc.status = IB_WC_SUCCESS;
+			wc.opcode = ib_qib_wc_opcode[wqe->wr.opcode];
+			wc.byte_len = wqe->length;
+			wc.qp = &qp->ibqp;
+			qib_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 0);
+		}
+		if (++qp->s_last >= qp->s_size)
+			qp->s_last = 0;
+	}
+	/*
+	 * If we were waiting for sends to complete before resending,
+	 * and they are now complete, restart sending.
+	 */
+	if (qp->s_flags & QIB_S_WAIT_PSN &&
+	    qib_cmp24(qp->s_sending_psn, qp->s_sending_hpsn) > 0) {
+		qp->s_flags &= ~QIB_S_WAIT_PSN;
+		qp->s_sending_psn = qp->s_psn;
+		qp->s_sending_hpsn = qp->s_psn - 1;
+		qib_schedule_send(qp);
+	}
+}
+
+static inline void update_last_psn(struct qib_qp *qp, u32 psn)
+{
+	qp->s_last_psn = psn;
+}
+
+/*
+ * Generate a SWQE completion.
+ * This is similar to qib_send_complete but has to check to be sure
+ * that the SGEs are not being referenced if the SWQE is being resent.
+ */
+static struct qib_swqe *do_rc_completion(struct qib_qp *qp,
+					 struct qib_swqe *wqe,
+					 struct qib_ibport *ibp)
+{
+	struct ib_wc wc;
+	unsigned i;
+
+	/*
+	 * Don't decrement refcount and don't generate a
+	 * completion if the SWQE is being resent until the send
+	 * is finished.
+	 */
+	if (qib_cmp24(wqe->lpsn, qp->s_sending_psn) < 0 ||
+	    qib_cmp24(qp->s_sending_psn, qp->s_sending_hpsn) > 0) {
+		for (i = 0; i < wqe->wr.num_sge; i++) {
+			struct qib_sge *sge = &wqe->sg_list[i];
+
+			atomic_dec(&sge->mr->refcount);
+		}
+		/* Post a send completion queue entry if requested. */
+		if (!(qp->s_flags & QIB_S_SIGNAL_REQ_WR) ||
+		    (wqe->wr.send_flags & IB_SEND_SIGNALED)) {
+			memset(&wc, 0, sizeof wc);
+			wc.wr_id = wqe->wr.wr_id;
+			wc.status = IB_WC_SUCCESS;
+			wc.opcode = ib_qib_wc_opcode[wqe->wr.opcode];
+			wc.byte_len = wqe->length;
+			wc.qp = &qp->ibqp;
+			qib_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 0);
+		}
+		if (++qp->s_last >= qp->s_size)
+			qp->s_last = 0;
+	} else
+		ibp->n_rc_delayed_comp++;
+
+	qp->s_retry = qp->s_retry_cnt;
+	update_last_psn(qp, wqe->lpsn);
+
+	/*
+	 * 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.
+	 */