IB/ipath: Change UD to queue work requests like RC & UC

The code to post UD sends tried to process work requests at the time
ib_post_send() is called without using a WQE queue.  This was fine as
long as HW resources were available for sending a packet.  This patch
changes UD to be handled more like RC and UC and shares more code.

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

7 files changed, 494 insertions, 611 deletions

diff --git a/drivers/infiniband/hw/ipath/ipath_qp.c b/drivers/infiniband/hw/ipath/ipath_qp.c
index 1324b35ff1f..a8c4a6b03d7 100644
--- a/drivers/infiniband/hw/ipath/ipath_qp.c
+++ b/drivers/infiniband/hw/ipath/ipath_qp.c
@@ -338,6 +338,7 @@ static void ipath_reset_qp(struct ipath_qp *qp)
 	qp->s_busy = 0;
 	qp->s_flags &= IPATH_S_SIGNAL_REQ_WR;
 	qp->s_hdrwords = 0;
+	qp->s_wqe = NULL;
 	qp->s_psn = 0;
 	qp->r_psn = 0;
 	qp->r_msn = 0;
@@ -751,6 +752,9 @@ struct ib_qp *ipath_create_qp(struct ib_pd *ibpd,
 	switch (init_attr->qp_type) {
 	case IB_QPT_UC:
 	case IB_QPT_RC:
+	case IB_QPT_UD:
+	case IB_QPT_SMI:
+	case IB_QPT_GSI:
 		sz = sizeof(struct ipath_sge) *
 			init_attr->cap.max_send_sge +
 			sizeof(struct ipath_swqe);
@@ -759,10 +763,6 @@ struct ib_qp *ipath_create_qp(struct ib_pd *ibpd,
 			ret = ERR_PTR(-ENOMEM);
 			goto bail;
 		}
-		/* FALLTHROUGH */
-	case IB_QPT_UD:
-	case IB_QPT_SMI:
-	case IB_QPT_GSI:
 		sz = sizeof(*qp);
 		if (init_attr->srq) {
 			struct ipath_srq *srq = to_isrq(init_attr->srq);
@@ -805,8 +805,7 @@ struct ib_qp *ipath_create_qp(struct ib_pd *ibpd,
 		spin_lock_init(&qp->r_rq.lock);
 		atomic_set(&qp->refcount, 0);
 		init_waitqueue_head(&qp->wait);
-		tasklet_init(&qp->s_task, ipath_do_ruc_send,
-			     (unsigned long)qp);
+		tasklet_init(&qp->s_task, ipath_do_send, (unsigned long)qp);
 		INIT_LIST_HEAD(&qp->piowait);
 		INIT_LIST_HEAD(&qp->timerwait);
 		qp->state = IB_QPS_RESET;
diff --git a/drivers/infiniband/hw/ipath/ipath_rc.c b/drivers/infiniband/hw/ipath/ipath_rc.c
index 46744ea2bab..53259daeb4f 100644
--- a/drivers/infiniband/hw/ipath/ipath_rc.c
+++ b/drivers/infiniband/hw/ipath/ipath_rc.c
@@ -81,9 +81,8 @@ static void ipath_init_restart(struct ipath_qp *qp, struct ipath_swqe *wqe)
  * Note that we are in the responder's side of the QP context.
  * Note the QP s_lock must be held.
  */
-static int ipath_make_rc_ack(struct ipath_qp *qp,
-			     struct ipath_other_headers *ohdr,
-			     u32 pmtu, u32 *bth0p, u32 *bth2p)
+static int ipath_make_rc_ack(struct ipath_ibdev *dev, struct ipath_qp *qp,
+			     struct ipath_other_headers *ohdr, u32 pmtu)
 {
 	struct ipath_ack_entry *e;
 	u32 hwords;
@@ -192,8 +191,7 @@ static int ipath_make_rc_ack(struct ipath_qp *qp,
 	}
 	qp->s_hdrwords = hwords;
 	qp->s_cur_size = len;
-	*bth0p = bth0 | (1 << 22); /* Set M bit */
-	*bth2p = bth2;
+	ipath_make_ruc_header(dev, qp, ohdr, bth0, bth2);
 	return 1;
 
 bail:
@@ -203,32 +201,39 @@ bail:
 /**
  * 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 and interrupts disabled.
  */
-int ipath_make_rc_req(struct ipath_qp *qp,
-		      struct ipath_other_headers *ohdr,
-		      u32 pmtu, u32 *bth0p, u32 *bth2p)
+int ipath_make_rc_req(struct ipath_qp *qp)
 {
 	struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
+	struct ipath_other_headers *ohdr;
 	struct ipath_sge_state *ss;
 	struct ipath_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;
+
+	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->r_head_ack_queue != qp->s_tail_ack_queue ||
 	     (qp->s_flags & IPATH_S_ACK_PENDING) ||
 	     qp->s_ack_state != OP(ACKNOWLEDGE)) &&
-	    ipath_make_rc_ack(qp, ohdr, pmtu, bth0p, bth2p))
+	    ipath_make_rc_ack(dev, qp, ohdr, pmtu))
 		goto done;
 
 	if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK) ||
@@ -560,13 +565,12 @@ int ipath_make_rc_req(struct ipath_qp *qp,
 	qp->s_hdrwords = hwords;
 	qp->s_cur_sge = ss;
 	qp->s_cur_size = len;
-	*bth0p = bth0 | (qp->s_state << 24);
-	*bth2p = bth2;
+	ipath_make_ruc_header(dev, qp, ohdr, bth0 | (qp->s_state << 24), bth2);
 done:
-	return 1;
-
+	ret = 1;
 bail:
-	return 0;
+	spin_unlock_irqrestore(&qp->s_lock, flags);
+	return ret;
 }
 
 /**
@@ -627,7 +631,7 @@ static void send_rc_ack(struct ipath_qp *qp)
 	/*
 	 * If we can send the ACK, clear the ACK state.
 	 */
-	if (ipath_verbs_send(dev->dd, hwords, (u32 *) &hdr, 0, NULL) == 0) {
+	if (ipath_verbs_send(qp, &hdr, hwords, NULL, 0) == 0) {
 		dev->n_unicast_xmit++;
 		goto done;
 	}
@@ -757,7 +761,9 @@ void ipath_restart_rc(struct ipath_qp *qp, u32 psn, struct ib_wc *wc)
 		wc->vendor_err = 0;
 		wc->byte_len = 0;
 		wc->qp = &qp->ibqp;
+		wc->imm_data = 0;
 		wc->src_qp = qp->remote_qpn;
+		wc->wc_flags = 0;
 		wc->pkey_index = 0;
 		wc->slid = qp->remote_ah_attr.dlid;
 		wc->sl = qp->remote_ah_attr.sl;
@@ -1041,7 +1047,9 @@ static int do_rc_ack(struct ipath_qp *qp, u32 aeth, u32 psn, int opcode,
 			wc.vendor_err = 0;
 			wc.byte_len = 0;
 			wc.qp = &qp->ibqp;
+			wc.imm_data = 0;
 			wc.src_qp = qp->remote_qpn;
+			wc.wc_flags = 0;
 			wc.pkey_index = 0;
 			wc.slid = qp->remote_ah_attr.dlid;
 			wc.sl = qp->remote_ah_attr.sl;
@@ -1454,6 +1462,19 @@ static inline int ipath_rc_rcv_error(struct ipath_ibdev *dev,
 			goto send_ack;
 		}
 		/*
+		 * Try to send a simple ACK to work around a Mellanox bug
+		 * which doesn't accept a RDMA read response or atomic
+		 * response as an ACK for earlier SENDs or RDMA writes.
+		 */
+		if (qp->r_head_ack_queue == qp->s_tail_ack_queue &&
+		    !(qp->s_flags & IPATH_S_ACK_PENDING) &&
+		    qp->s_ack_state == OP(ACKNOWLEDGE)) {
+			spin_unlock_irqrestore(&qp->s_lock, flags);
+			qp->r_nak_state = 0;
+			qp->r_ack_psn = qp->s_ack_queue[i].psn - 1;
+			goto send_ack;
+		}
+		/*
 		 * Resend the RDMA read or atomic op which
 		 * ACKs this duplicate request.
 		 */
diff --git a/drivers/infiniband/hw/ipath/ipath_ruc.c b/drivers/infiniband/hw/ipath/ipath_ruc.c
index c69c2523944..4b6b7ee8e5c 100644
--- a/drivers/infiniband/hw/ipath/ipath_ruc.c
+++ b/drivers/infiniband/hw/ipath/ipath_ruc.c
@@ -31,6 +31,8 @@
  * SOFTWARE.
  */
 
+#include <linux/spinlock.h>
+
 #include "ipath_verbs.h"
 #include "ipath_kernel.h"
 
@@ -106,27 +108,30 @@ void ipath_insert_rnr_queue(struct ipath_qp *qp)
 	spin_unlock_irqrestore(&dev->pending_lock, flags);
 }
 
-static int init_sge(struct ipath_qp *qp, struct ipath_rwqe *wqe)
+/**
+ * ipath_init_sge - Validate a RWQE and fill in the SGE state
+ * @qp: the QP
+ *
+ * Return 1 if OK.
+ */
+int ipath_init_sge(struct ipath_qp *qp, struct ipath_rwqe *wqe,
+		   u32 *lengthp, struct ipath_sge_state *ss)
 {
-	int user = to_ipd(qp->ibqp.pd)->user;
 	int i, j, ret;
 	struct ib_wc wc;
 
-	qp->r_len = 0;
+	*lengthp = 0;
 	for (i = j = 0; i < wqe->num_sge; i++) {
 		if (wqe->sg_list[i].length == 0)
 			continue;
 		/* Check LKEY */
-		if ((user && wqe->sg_list[i].lkey == 0) ||
-		    !ipath_lkey_ok(qp, &qp->r_sg_list[j], &wqe->sg_list[i],
-				   IB_ACCESS_LOCAL_WRITE))
+		if (!ipath_lkey_ok(qp, j ? &ss->sg_list[j - 1] : &ss->sge,
+				   &wqe->sg_list[i], IB_ACCESS_LOCAL_WRITE))
 			goto bad_lkey;
-		qp->r_len += wqe->sg_list[i].length;
+		*lengthp += wqe->sg_list[i].length;
 		j++;
 	}
-	qp->r_sge.sge = qp->r_sg_list[0];
-	qp->r_sge.sg_list = qp->r_sg_list + 1;
-	qp->r_sge.num_sge = j;
+	ss->num_sge = j;
 	ret = 1;
 	goto bail;
 
@@ -172,6 +177,8 @@ int ipath_get_rwqe(struct ipath_qp *qp, int wr_id_only)
 	u32 tail;
 	int ret;
 
+	qp->r_sge.sg_list = qp->r_sg_list;
+
 	if (qp->ibqp.srq) {
 		srq = to_isrq(qp->ibqp.srq);
 		handler = srq->ibsrq.event_handler;
@@ -199,7 +206,8 @@ int ipath_get_rwqe(struct ipath_qp *qp, int wr_id_only)
 		wqe = get_rwqe_ptr(rq, tail);
 		if (++tail >= rq->size)
 			tail = 0;
-	} while (!wr_id_only && !init_sge(qp, wqe));
+	} while (!wr_id_only && !ipath_init_sge(qp, wqe, &qp->r_len,
+						&qp->r_sge));
 	qp->r_wr_id = wqe->wr_id;
 	wq->tail = tail;
 
@@ -239,9 +247,9 @@ bail:
 
 /**
  * ipath_ruc_loopback - handle UC and RC lookback requests
- * @sqp: the loopback QP
+ * @sqp: the sending QP
  *
- * This is called from ipath_do_uc_send() or ipath_do_rc_send() to
+ * This is called from ipath_do_send() to
  * forward a WQE addressed to the same HCA.
  * Note that although we are single threaded due to the tasklet, we still
  * have to protect against post_send().  We don't have to worry about
@@ -450,40 +458,18 @@ again:
 	wc.byte_len = wqe->length;
 	wc.qp = &qp->ibqp;
 	wc.src_qp = qp->remote_qpn;
-	/* XXX do we know which pkey matched? Only needed for GSI. */
 	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 = 1;
 	/* Signal completion event if the solicited bit is set. */
 	ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
 		       wqe->wr.send_flags & IB_SEND_SOLICITED);
 
 send_comp:
 	sqp->s_rnr_retry = sqp->s_rnr_retry_cnt;
-
-	if (!(sqp->s_flags & IPATH_S_SIGNAL_REQ_WR) ||
-	    (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 = &sqp->ibqp;
-		wc.src_qp = 0;
-		wc.pkey_index = 0;
-		wc.slid = 0;
-		wc.sl = 0;
-		wc.dlid_path_bits = 0;
-		wc.port_num = 0;
-		ipath_cq_enter(to_icq(sqp->ibqp.send_cq), &wc, 0);
-	}
-
-	/* Update s_last now that we are finished with the SWQE */
-	spin_lock_irqsave(&sqp->s_lock, flags);
-	if (++sqp->s_last >= sqp->s_size)
-		sqp->s_last = 0;
-	spin_unlock_irqrestore(&sqp->s_lock, flags);
+	ipath_send_complete(sqp, wqe, IB_WC_SUCCESS);
 	goto again;
 
 done:
@@ -491,13 +477,11 @@ done:
 		wake_up(&qp->wait);
 }
 
-static int want_buffer(struct ipath_devdata *dd)
+static void want_buffer(struct ipath_devdata *dd)
 {
 	set_bit(IPATH_S_PIOINTBUFAVAIL, &dd->ipath_sendctrl);
 	ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
 			 dd->ipath_sendctrl);
-
-	return 0;
 }
 
 /**
@@ -507,14 +491,11 @@ static int want_buffer(struct ipath_devdata *dd)
  *
  * Called when we run out of PIO buffers.
  */
-static void ipath_no_bufs_available(struct ipath_qp *qp, struct ipath_ibdev *dev)
+static void ipath_no_bufs_available(struct ipath_qp *qp,
+				    struct ipath_ibdev *dev)
 {
 	unsigned long flags;
 
-	spin_lock_irqsave(&dev->pending_lock, flags);
-	if (list_empty(&qp->piowait))
-		list_add_tail(&qp->piowait, &dev->piowait);
-	spin_unlock_irqrestore(&dev->pending_lock, flags);
 	/*
 	 * Note that as soon as want_buffer() is called and
 	 * possibly before it returns, ipath_ib_piobufavail()
@@ -524,101 +505,14 @@ static void ipath_no_bufs_available(struct ipath_qp *qp, struct ipath_ibdev *dev
 	 * We leave the busy flag set so that another post send doesn't
 	 * try to put the same QP on the piowait list again.
 	 */
+	spin_lock_irqsave(&dev->pending_lock, flags);
+	list_add_tail(&qp->piowait, &dev->piowait);
+	spin_unlock_irqrestore(&dev->pending_lock, flags);
 	want_buffer(dev->dd);
 	dev->n_piowait++;
 }
 
 /**
- * ipath_post_ruc_send - post RC and UC sends
- * @qp: the QP to post on
- * @wr: the work request to send
- */
-int ipath_post_ruc_send(struct ipath_qp *qp, struct ib_send_wr *wr)
-{
-	struct ipath_swqe *wqe;
-	unsigned long flags;
-	u32 next;
-	int i, j;
-	int acc;
-	int ret;
-
-	/*
-	 * Don't allow RDMA reads or atomic operations on UC or
-	 * undefined operations.
-	 * Make sure buffer is large enough to hold the result for atomics.
-	 */
-	if (qp->ibqp.qp_type == IB_QPT_UC) {
-		if ((unsigned) wr->opcode >= IB_WR_RDMA_READ) {
-			ret = -EINVAL;
-			goto bail;
-		}
-	} else if ((unsigned) wr->opcode > IB_WR_ATOMIC_FETCH_AND_ADD) {
-		ret = -EINVAL;
-		goto bail;
-	} else if (wr->opcode >= IB_WR_ATOMIC_CMP_AND_SWP &&
-		   (wr->num_sge == 0 ||
-		    wr->sg_list[0].length < sizeof(u64) ||
-		    wr->sg_list[0].addr & (sizeof(u64) - 1))) {
-		ret = -EINVAL;
-		goto bail;
-	} else if (wr->opcode >= IB_WR_RDMA_READ && !qp->s_max_rd_atomic) {
-		ret = -EINVAL;
-		goto bail;
-	}
-	/* IB spec says that num_sge == 0 is OK. */
-	if (wr->num_sge > qp->s_max_sge) {
-		ret = -ENOMEM;
-		goto bail;
-	}
-	spin_lock_irqsave(&qp->s_lock, flags);
-	next = qp->s_head + 1;
-	if (next >= qp->s_size)
-		next = 0;
-	if (next == qp->s_last) {
-		spin_unlock_irqrestore(&qp->s_lock, flags);
-		ret = -EINVAL;
-		goto bail;
-	}
-
-	wqe = get_swqe_ptr(qp, qp->s_head);
-	wqe->wr = *wr;
-	wqe->ssn = qp->s_ssn++;
-	wqe->sg_list[0].mr = NULL;
-	wqe->sg_list[0].vaddr = NULL;
-	wqe->sg_list[0].length = 0;
-	wqe->sg_list[0].sge_length = 0;
-	wqe->length = 0;
-	acc = wr->opcode >= IB_WR_RDMA_READ ? IB_ACCESS_LOCAL_WRITE : 0;
-	for (i = 0, j = 0; i < wr->num_sge; i++) {
-		if (to_ipd(qp->ibqp.pd)->user && wr->sg_list[i].lkey == 0) {
-			spin_unlock_irqrestore(&qp->s_lock, flags);
-			ret = -EINVAL;
-			goto bail;
-		}
-		if (wr->sg_list[i].length == 0)
-			continue;
-		if (!ipath_lkey_ok(qp, &wqe->sg_list[j], &wr->sg_list[i],
-				   acc)) {
-			spin_unlock_irqrestore(&qp->s_lock, flags);
-			ret = -EINVAL;
-			goto bail;
-		}
-		wqe->length += wr->sg_list[i].length;
-		j++;
-	}
-	wqe->wr.num_sge = j;
-	qp->s_head = next;
-	spin_unlock_irqrestore(&qp->s_lock, flags);
-
-	ipath_do_ruc_send((unsigned long) qp);
-
-	ret = 0;
-
-bail:
-	return ret;
-}
-
-/**
  * ipath_make_grh - construct a GRH header
  * @dev: a pointer to the ipath device
  * @hdr: a pointer to the GRH header being constructed
@@ -648,39 +542,66 @@ u32 ipath_make_grh(struct ipath_ibdev *dev, struct ib_grh *hdr,
 	return sizeof(struct ib_grh) / sizeof(u32);
 }
 
+void ipath_make_ruc_header(struct ipath_ibdev *dev, struct ipath_qp *qp,
+			   struct ipath_other_headers *ohdr,
+			   u32 bth0, u32 bth2)
+{
+	u16 lrh0;
+	u32 nwords;
+	u32 extra_bytes;
+
+	/* Construct the header. */
+	extra_bytes = -qp->s_cur_size & 3;
+	nwords = (qp->s_cur_size + extra_bytes) >> 2;
+	lrh0 = IPATH_LRH_BTH;
+	if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
+		qp->s_hdrwords += ipath_make_grh(dev, &qp->s_hdr.u.l.grh,
+						 &qp->remote_ah_attr.grh,
+						 qp->s_hdrwords, nwords);
+		lrh0 = IPATH_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(dev->dd->ipath_lid);
+	bth0 |= ipath_get_pkey(dev->dd, qp->s_pkey_index);
+	bth0 |= extra_bytes << 20;
+	ohdr->bth[0] = cpu_to_be32(bth0 | (1 << 22));
+	ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
+	ohdr->bth[2] = cpu_to_be32(bth2);
+}
+
 /**
- * ipath_do_ruc_send - perform a send on an RC or UC QP
+ * ipath_do_send - perform a send on a 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.
+ * Otherwise, two threads could send packets out of order.
  */
-void ipath_do_ruc_send(unsigned long data)
+void ipath_do_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;
+	int (*make_req)(struct ipath_qp *qp);
 
 	if (test_and_set_bit(IPATH_S_BUSY, &qp->s_busy))
 		goto bail;
 
-	if (unlikely(qp->remote_ah_attr.dlid == dev->dd->ipath_lid)) {
+	if ((qp->ibqp.qp_type == IB_QPT_RC ||
+	     qp->ibqp.qp_type == IB_QPT_UC) &&
+	    qp->remote_ah_attr.dlid == dev->dd->ipath_lid) {
 		ipath_ruc_loopback(qp);
 		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;
+	if (qp->ibqp.qp_type == IB_QPT_RC)
+	       make_req = ipath_make_rc_req;
+	else if (qp->ibqp.qp_type == IB_QPT_UC)
+	       make_req = ipath_make_uc_req;
+	else
+	       make_req = ipath_make_ud_req;
 
 again:
 	/* Check for a constructed packet to be sent. */
@@ -689,9 +610,8 @@ again:
 		 * If no PIO bufs are available, return.  An interrupt will
 		 * call ipath_ib_piobufavail() when one is available.
 		 */
-		if (ipath_verbs_send(dev->dd, qp->s_hdrwords,
-				     (u32 *) &qp->s_hdr, qp->s_cur_size,
-				     qp->s_cur_sge)) {
+		if (ipath_verbs_send(qp, &qp->s_hdr, qp->s_hdrwords,
+				     qp->s_cur_sge, qp->s_cur_size)) {
 			ipath_no_bufs_available(qp, dev);
 			goto bail;
 		}
@@ -700,54 +620,42 @@ again:
 		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);
-
-	if (!((qp->ibqp.qp_type == IB_QPT_RC) ?
-	       ipath_make_rc_req(qp, ohdr, pmtu, &bth0, &bth2) :
-	       ipath_make_uc_req(qp, ohdr, pmtu, &bth0, &bth2))) {
-		/*
-		 * Clear the busy bit before unlocking to avoid races with
-		 * adding new work queue items and then failing to process
-		 * them.
-		 */
-		clear_bit(IPATH_S_BUSY, &qp->s_busy);
-		spin_unlock_irqrestore(&qp->s_lock, flags);
-		goto bail;
-	}
+	if (make_req(qp))
+		goto again;
+clear:
+	clear_bit(IPATH_S_BUSY, &qp->s_busy);
+bail:;
+}
 
-	spin_unlock_irqrestore(&qp->s_lock, flags);
+void ipath_send_complete(struct ipath_qp *qp, struct ipath_swqe *wqe,
+			 enum ib_wc_status status)
+{
+	u32 last = qp->s_last;
 
-	/* Construct the header. */
-	extra_bytes = (4 - qp->s_cur_size) & 3;
-	nwords = (qp->s_cur_size + extra_bytes) >> 2;
-	lrh0 = IPATH_LRH_BTH;
-	if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
-		qp->s_hdrwords += ipath_make_grh(dev, &qp->s_hdr.u.l.grh,
-						 &qp->remote_ah_attr.grh,
-						 qp->s_hdrwords, nwords);
-		lrh0 = IPATH_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(dev->dd->ipath_lid);
-	bth0 |= ipath_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);
+	if (++last == qp->s_size)
+		last = 0;
+	qp->s_last = last;
 
-	/* Check for more work to do. */
-	goto again;
+	/* See ch. 11.2.4.1 and 10.7.3.1 */
+	if (!(qp->s_flags & IPATH_S_SIGNAL_REQ_WR) ||
+	    (wqe->wr.send_flags & IB_SEND_SIGNALED) ||
+	    status != IB_WC_SUCCESS) {
+		struct ib_wc wc;
 
-clear:
-	clear_bit(IPATH_S_BUSY, &qp->s_busy);
-bail:
-	return;
+		wc.wr_id = wqe->wr.wr_id;
+		wc.status = status;
+		wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
+		wc.vendor_err = 0;
+		wc.byte_len = wqe->length;
+		wc.imm_data = 0;
+		wc.qp = &qp->ibqp;
+		wc.src_qp = 0;
+		wc.wc_flags = 0;
+		wc.pkey_index = 0;
+		wc.slid = 0;
+		wc.sl = 0;
+		wc.dlid_path_bits = 0;
+		wc.port_num = 0;
+		ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 0);
+	}
 }
diff --git a/drivers/infiniband/hw/ipath/ipath_uc.c b/drivers/infiniband/hw/ipath/ipath_uc.c
index 8380fbc50d2..767beb903c2 100644
--- a/drivers/infiniband/hw/ipath/ipath_uc.c
+++ b/drivers/infiniband/hw/ipath/ipath_uc.c
@@ -37,72 +37,40 @@
 /* cut down ridiculously long IB macro names */
 #define OP(x) IB_OPCODE_UC_##x
 
-static void complete_last_send(struct ipath_qp *qp, struct ipath_swqe *wqe,
-			       struct ib_wc *wc)
-{
-	if (++qp->s_last == qp->s_size)
-		qp->s_last = 0;
-	if (!(qp->s_flags & IPATH_S_SIGNAL_REQ_WR) ||
-	    (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 = &qp->ibqp;
-		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);
-	}
-}
-
 /**
  * ipath_make_uc_req - construct a request packet (SEND, RDMA write)
  * @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 and interrupts disabled.
  */
-int ipath_make_uc_req(struct ipath_qp *qp,
-		      struct ipath_other_headers *ohdr,
-		      u32 pmtu, u32 *bth0p, u32 *bth2p)
+int ipath_make_uc_req(struct ipath_qp *qp)
 {
+	struct ipath_other_headers *ohdr;
 	struct ipath_swqe *wqe;
 	u32 hwords;
 	u32 bth0;
 	u32 len;
-	struct ib_wc wc;
+	u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
+	int ret = 0;
 
 	if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK))
 		goto done;
 
+	ohdr = &qp->s_hdr.u.oth;
+	if (qp->remote_ah_attr.ah_flags & IB_AH_GRH)
+		ohdr = &qp->s_hdr.u.l.oth;
+
 	/* header size in 32-bit words LRH+BTH = (8+12)/4. */
 	hwords = 5;
 	bth0 = 1 << 22; /* Set M bit */
 
 	/* Get the next send request. */
-	wqe = get_swqe_ptr(qp, qp->s_last);
+	wqe = get_swqe_ptr(qp, qp->s_cur);
+	qp->s_wqe = NULL;
 	switch (qp->s_state) {
 	default:
-		/*
-		 * Signal the completion of the last send
-		 * (if there is one).
-		 */
-		if (qp->s_last != qp->s_tail) {
-			complete_last_send(qp, wqe, &wc);
-			wqe = get_swqe_ptr(qp, qp->s_last);
-		}
-
 		/* Check if send work queue is empty. */
-		if (qp->s_tail == qp->s_head)
+		if (qp->s_cur == qp->s_head)
 			goto done;
 		/*
 		 * Start a new request.
@@ -131,6 +99,9 @@ int ipath_make_uc_req(struct ipath_qp *qp,
 			}
 			if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 				bth0 |= 1 << 23;
+			qp->s_wqe = wqe;
+			if (++qp->s_cur >= qp->s_size)
+				qp->s_cur = 0;
 			break;
 
 		case IB_WR_RDMA_WRITE:
@@ -157,13 +128,14 @@ int ipath_make_uc_req(struct ipath_qp *qp,
 				if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 					bth0 |= 1 << 23;
 			}
+			qp->s_wqe = wqe;
+			if (++qp->s_cur >= qp->s_size)
+				qp->s_cur = 0;
 			break;
 
 		default:
 			goto done;
 		}
-		if (++qp->s_tail >= qp->s_size)
-			qp->s_tail = 0;
 		break;
 
 	case OP(SEND_FIRST):
@@ -185,6 +157,9 @@ int ipath_make_uc_req(struct ipath_qp *qp,
 		}
 		if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 			bth0 |= 1 << 23;
+		qp->s_wqe = wqe;
+		if (++qp->s_cur >= qp->s_size)
+			qp->s_cur = 0;
 		break;
 
 	case OP(RDMA_WRITE_FIRST):
@@ -207,18 +182,22 @@ int ipath_make_uc_req(struct ipath_qp *qp,
 			if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 				bth0 |= 1 << 23;
 		}
+		qp->s_wqe = wqe;
+		if (++qp->s_cur >= qp->s_size)
+			qp->s_cur = 0;
 		break;
 	}
 	qp->s_len -= len;
 	qp->s_hdrwords = hwords;
 	qp->s_cur_sge = &qp->s_sge;
 	qp->s_cur_size = len;
-	*bth0p = bth0 | (qp->s_state << 24);
-	*bth2p = qp->s_next_psn++ & IPATH_PSN_MASK;
-	return 1;
+	ipath_make_ruc_header(to_idev(qp->ibqp.device),
+			      qp, ohdr, bth0 | (qp->s_state << 24),
+			      qp->s_next_psn++ & IPATH_PSN_MASK);
+	ret = 1;
 
 done:
-	return 0;
+	return ret;
 }
 
 /**
diff --git a/drivers/infiniband/hw/ipath/ipath_ud.c b/drivers/infiniband/hw/ipath/ipath_ud.c
index f9a3338a5fb..34c4a0a5be3 100644
--- a/drivers/infiniband/hw/ipath/ipath_ud.c
+++ b/drivers/infiniband/hw/ipath/ipath_ud.c
@@ -36,68 +36,17 @@
 #include "ipath_verbs.h"
 #include "ipath_kernel.h"
 
-static int init_sge(struct ipath_qp *qp, struct ipath_rwqe *wqe,
-		    u32 *lengthp, struct ipath_sge_state *ss)
-{
-	int user = to_ipd(qp->ibqp.pd)->user;
-	int i, j, ret;
-	struct ib_wc wc;
-
-	*lengthp = 0;
-	for (i = j = 0; i < wqe->num_sge; i++) {
-		if (wqe->sg_list[i].length == 0)
-			continue;
-		/* Check LKEY */
-		if ((user && wqe->sg_list[i].lkey == 0) ||
-		    !ipath_lkey_ok(qp, j ? &ss->sg_list[j - 1] : &ss->sge,
-				   &wqe->sg_list[i], IB_ACCESS_LOCAL_WRITE))
-			goto bad_lkey;
-		*lengthp += wqe->sg_list[i].length;
-		j++;
-	}
-	ss->num_sge = j;
-	ret = 1;
-	goto bail;
-
-bad_lkey:
-	wc.wr_id = wqe->wr_id;
-	wc.status = IB_WC_LOC_PROT_ERR;
-	wc.opcode = IB_WC_RECV;
-	wc.vendor_err = 0;
-	wc.byte_len = 0;
-	wc.imm_data = 0;
-	wc.qp = &qp->ibqp;
-	wc.src_qp = 0;
-	wc.wc_flags = 0;
-	wc.pkey_index = 0;
-	wc.slid = 0;
-	wc.sl = 0;
-	wc.dlid_path_bits = 0;
-	wc.port_num = 0;
-	/* Signal solicited completion event. */
-	ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
-	ret = 0;
-bail:
-	return ret;
-}
-
 /**
  * ipath_ud_loopback - handle send on loopback QPs
- * @sqp: the QP
- * @ss: the SGE state
- * @length: the length of the data to send
- * @wr: the work request
- * @wc: the work completion entry
+ * @sqp: the sending QP
+ * @swqe: the send work request
  *
- * This is called from ipath_post_ud_send() to forward a WQE addressed
+ * This is called from ipath_make_ud_req() to forward a WQE addressed
  * to the same HCA.
  * Note that the receive interrupt handler may be calling ipath_ud_rcv()
  * while this is being called.
  */
-static void ipath_ud_loopback(struct ipath_qp *sqp,
-			      struct ipath_sge_state *ss,
-			      u32 length, struct ib_send_wr *wr,
-			      struct ib_wc *wc)
+static void ipath_ud_loopback(struct ipath_qp *sqp, struct ipath_swqe *swqe)
 {
 	struct ipath_ibdev *dev = to_idev(sqp->ibqp.device);
 	struct ipath_qp *qp;
@@ -110,12 +59,18 @@ static void ipath_ud_loopback(struct ipath_qp *sqp,
 	struct ipath_rwq *wq;
 	struct ipath_rwqe *wqe;
 	void (*handler)(struct ib_event *, void *);
+	struct ib_wc wc;
 	u32 tail;
 	u32 rlen;
+	u32 length;
 
-	qp = ipath_lookup_qpn(&dev->qp_table, wr->wr.ud.remote_qpn);
-	if (!qp)
-		return;
+	qp = ipath_lookup_qpn(&dev->qp_table, swqe->wr.wr.ud.remote_qpn);
+	if (!qp) {
+		dev->n_pkt_drops++;
+		goto send_comp;
+	}
+
+	rsge.sg_list = NULL;
 
 	/*
 	 * Check that the qkey matches (except for QP0, see 9.6.1.4.1).
@@ -123,39 +78,34 @@ static void ipath_ud_loopback(struct ipath_qp *sqp,
 	 * qkey from the QP context instead of the WR (see 10.2.5).
 	 */
 	if (unlikely(qp->ibqp.qp_num &&
-		     ((int) wr->wr.ud.remote_qkey < 0
-		      ? qp->qkey : wr->wr.ud.remote_qkey) != qp->qkey)) {
+		     ((int) swqe->wr.wr.ud.remote_qkey < 0 ?
+		      sqp->qkey : swqe->wr.wr.ud.remote_qkey) != qp->qkey)) {
 		/* XXX OK to lose a count once in a while. */
 		dev->qkey_violations++;
 		dev->n_pkt_drops++;
-		goto done;
+		goto drop;
 	}
 
 	/*
 	 * A GRH is expected to preceed the data even if not
 	 * present on the wire.
 	 */
-	wc->byte_len = length + sizeof(struct ib_grh);
+	length = swqe->length;
+	wc.byte_len = length + sizeof(struct ib_grh);
 
-	if (wr->opcode == IB_WR_SEND_WITH_IMM) {
-		wc->wc_flags = IB_WC_WITH_IMM;
-		wc->imm_data = wr->imm_data;
+	if (swqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
+		wc.wc_flags = IB_WC_WITH_IMM;
+		wc.imm_data = swqe->wr.imm_data;
 	} else {
-		wc->wc_flags = 0;
-		wc->imm_data = 0;
+		wc.wc_flags = 0;
+		wc.imm_data = 0;
 	}
 
-	if (wr->num_sge > 1) {
-		rsge.sg_list = kmalloc((wr->num_sge - 1) *
-					sizeof(struct ipath_sge),
-				       GFP_ATOMIC);
-	} else
-		rsge.sg_list = NULL;
-
 	/*
-	 * Get the next work request entry to find where to put the data.
-	 * Note that it is safe to drop the lock after changing rq->tail
-	 * since ipath_post_receive() won't fill the empty slot.
+	 * This would be a lot simpler if we could call ipath_get_rwqe()
+	 * but that uses state that the receive interrupt handler uses
+	 * so we would need to lock out receive interrupts while doing
+	 * local loopback.
 	 */
 	if (qp->ibqp.srq) {
 		srq = to_isrq(qp->ibqp.srq);
@@ -167,32 +117,53 @@ static void ipath_ud_loopback(struct ipath_qp *sqp,
 		rq = &qp->r_rq;
 	}
 
+	if (rq->max_sge > 1) {
+		/*
+		 * XXX We could use GFP_KERNEL if ipath_do_send()
+		 * was always called from the tasklet instead of
+		 * from ipath_post_send().
+		 */
+		rsge.sg_list = kmalloc((rq->max_sge - 1) *
+					sizeof(struct ipath_sge),
+				       GFP_ATOMIC);
+		if (!rsge.sg_list) {
+			dev->n_pkt_drops++;
+			goto drop;
+		}
+	}
+
+	/*
+	 * Get the next work request entry to find where to put the data.
+	 * Note that it is safe to drop the lock after changing rq->tail
+	 * since ipath_post_receive() won't fill the empty slot.
+	 */
 	spin_lock_irqsave(&rq->lock, flags);
 	wq = rq->wq;
 	tail = wq->tail;
-	while (1) {
-		if (unlikely(tail == wq->head)) {
-			spin_unlock_irqrestore(&rq->lock, flags);
-			dev->n_pkt_drops++;
-			goto bail_sge;
-		}
-		/* Make sure entry is read after head index is read. */
-		smp_rmb();
-		wqe = get_rwqe_ptr(rq, tail);
-		if (++tail >= rq->size)
-			tail = 0;
-		if (init_sge(qp, wqe, &rlen, &rsge))
-			break;
-		wq->tail = tail;
+	/* Validate tail before using it since it is user writable. */
+	if (tail >= rq->size)
+		tail = 0;
+	if (unlikely(tail == wq->head)) {
+		spin_unlock_irqrestore(&rq->lock, flags);
+		dev->n_pkt_drops++;
+		goto drop;
+	}
+	wqe = get_rwqe_ptr(rq, tail);
+	if (!ipath_init_sge(qp, wqe, &rlen, &rsge)) {
+		spin_unlock_irqrestore(&rq->lock, flags);
+		dev->n_pkt_drops++;
+		goto drop;
 	}
 	/* Silently drop packets which are too big. */
-	if (wc->byte_len > rlen) {
+	if (wc.byte_len > rlen) {
 		spin_unlock_irqrestore(&rq->lock, flags);
 		dev->n_pkt_drops++;
-		goto bail_sge;
+		goto drop;
 	}
+	if (++tail >= rq->size)
+		tail = 0;
 	wq->tail = tail;
-	wc->wr_id = wqe->wr_id;
+	wc.wr_id = wqe->wr_id;
 	if (handler) {
 		u32 n;
 
@@ -221,13 +192,13 @@ static void ipath_ud_loopback(struct ipath_qp *sqp,
 	} else
 		spin_unlock_irqrestore(&rq->lock, flags);
 
-	ah_attr = &to_iah(wr->wr.ud.ah)->attr;
+	ah_attr = &to_iah(swqe->wr.wr.ud.ah)->attr;
 	if (ah_attr->ah_flags & IB_AH_GRH) {
 		ipath_copy_sge(&rsge, &ah_attr->grh, sizeof(struct ib_grh));
-		wc->wc_flags |= IB_WC_GRH;
+		wc.wc_flags |= IB_WC_GRH;
 	} else
 		ipath_skip_sge(&rsge, sizeof(struct ib_grh));
-	sge = &ss->sge;
+	sge = swqe->sg_list;
 	while (length) {
 		u32 len = sge->length;
 
@@ -241,8