1 files changed, 311 insertions, 313 deletions

diff --git a/drivers/infiniband/hw/ipath/ipath_ruc.c b/drivers/infiniband/hw/ipath/ipath_ruc.c
index 5c1da2d25e0..1f95bbaf760 100644
--- a/drivers/infiniband/hw/ipath/ipath_ruc.c
+++ b/drivers/infiniband/hw/ipath/ipath_ruc.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2006 QLogic, Inc. All rights reserved.
+ * Copyright (c) 2006, 2007, 2008 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
@@ -31,6 +31,9 @@
  * SOFTWARE.
  */
 
+#include <linux/sched.h>
+#include <linux/spinlock.h>
+
 #include "ipath_verbs.h"
 #include "ipath_kernel.h"
 
@@ -76,6 +79,7 @@ const u32 ib_ipath_rnr_table[32] = {
  * ipath_insert_rnr_queue - put QP on the RNR timeout list for the device
  * @qp: the QP
  *
+ * Called with the QP s_lock held and interrupts disabled.
  * XXX Use a simple list for now.  We might need a priority
  * queue if we have lots of QPs waiting for RNR timeouts
  * but that should be rare.
@@ -83,9 +87,9 @@ const u32 ib_ipath_rnr_table[32] = {
 void ipath_insert_rnr_queue(struct ipath_qp *qp)
 {
 	struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
-	unsigned long flags;
 
-	spin_lock_irqsave(&dev->pending_lock, flags);
+	/* We already did a spin_lock_irqsave(), so just use spin_lock */
+	spin_lock(&dev->pending_lock);
 	if (list_empty(&dev->rnrwait))
 		list_add(&qp->timerwait, &dev->rnrwait);
 	else {
@@ -96,57 +100,53 @@ void ipath_insert_rnr_queue(struct ipath_qp *qp)
 		while (qp->s_rnr_timeout >= nqp->s_rnr_timeout) {
 			qp->s_rnr_timeout -= nqp->s_rnr_timeout;
 			l = l->next;
-			if (l->next == &dev->rnrwait)
+			if (l->next == &dev->rnrwait) {
+				nqp = NULL;
 				break;
+			}
 			nqp = list_entry(l->next, struct ipath_qp,
 					 timerwait);
 		}
+		if (nqp)
+			nqp->s_rnr_timeout -= qp->s_rnr_timeout;
 		list_add(&qp->timerwait, l);
 	}
-	spin_unlock_irqrestore(&dev->pending_lock, flags);
+	spin_unlock(&dev->pending_lock);
 }
 
-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)
 {
-	struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
-	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(&dev->lk_table,
-				   &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;
 
 bad_lkey:
+	memset(&wc, 0, sizeof(wc));
 	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_num = qp->ibqp.qp_num;
-	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;
+	wc.qp = &qp->ibqp;
 	/* Signal solicited completion event. */
 	ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
 	ret = 0;
@@ -157,7 +157,7 @@ bail:
 /**
  * ipath_get_rwqe - copy the next RWQE into the QP's RWQE
  * @qp: the QP
- * @wr_id_only: update wr_id only, not SGEs
+ * @wr_id_only: update qp->r_wr_id only, not qp->r_sge
  *
  * Return 0 if no RWQE is available, otherwise return 1.
  *
@@ -185,6 +185,11 @@ int ipath_get_rwqe(struct ipath_qp *qp, int wr_id_only)
 	}
 
 	spin_lock_irqsave(&rq->lock, flags);
+	if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_RECV_OK)) {
+		ret = 0;
+		goto unlock;
+	}
+
 	wq = rq->wq;
 	tail = wq->tail;
 	/* Validate tail before using it since it is user writable. */
@@ -192,18 +197,23 @@ int ipath_get_rwqe(struct ipath_qp *qp, int wr_id_only)
 		tail = 0;
 	do {
 		if (unlikely(tail == wq->head)) {
-			spin_unlock_irqrestore(&rq->lock, flags);
 			ret = 0;
-			goto bail;
+			goto unlock;
 		}
+		/* Make sure entry is read after head index is read. */
+		smp_rmb();
 		wqe = get_rwqe_ptr(rq, tail);
 		if (++tail >= rq->size)
 			tail = 0;
-	} while (!wr_id_only && !init_sge(qp, wqe));
+		if (wr_id_only)
+			break;
+		qp->r_sge.sg_list = qp->r_sg_list;
+	} while (!ipath_init_sge(qp, wqe, &qp->r_len, &qp->r_sge));
 	qp->r_wr_id = wqe->wr_id;
 	wq->tail = tail;
 
 	ret = 1;
+	set_bit(IPATH_R_WRID_VALID, &qp->r_aflags);
 	if (handler) {
 		u32 n;
 
@@ -230,17 +240,17 @@ int ipath_get_rwqe(struct ipath_qp *qp, int wr_id_only)
 			goto bail;
 		}
 	}
+unlock:
 	spin_unlock_irqrestore(&rq->lock, flags);
-
 bail:
 	return ret;
 }
 
 /**
  * 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
@@ -256,37 +266,64 @@ static void ipath_ruc_loopback(struct ipath_qp *sqp)
 	unsigned long flags;
 	struct ib_wc wc;
 	u64 sdata;
+	atomic64_t *maddr;
+	enum ib_wc_status send_status;
 
+	/*
+	 * Note that we check the responder QP state after
+	 * checking the requester's state.
+	 */
 	qp = ipath_lookup_qpn(&dev->qp_table, sqp->remote_qpn);
-	if (!qp) {
-		dev->n_pkt_drops++;
-		return;
-	}
 
-again:
 	spin_lock_irqsave(&sqp->s_lock, flags);
 
-	if (!(ib_ipath_state_ops[sqp->state] & IPATH_PROCESS_SEND_OK)) {
-		spin_unlock_irqrestore(&sqp->s_lock, flags);
-		goto done;
-	}
+	/* Return if we are already busy processing a work request. */
+	if ((sqp->s_flags & (IPATH_S_BUSY | IPATH_S_ANY_WAIT)) ||
+	    !(ib_ipath_state_ops[sqp->state] & IPATH_PROCESS_OR_FLUSH_SEND))
+		goto unlock;
 
-	/* Get the next send request. */
-	if (sqp->s_last == sqp->s_head) {
-		/* Send work queue is empty. */
-		spin_unlock_irqrestore(&sqp->s_lock, flags);
-		goto done;
+	sqp->s_flags |= IPATH_S_BUSY;
+
+again:
+	if (sqp->s_last == sqp->s_head)
+		goto clr_busy;
+	wqe = get_swqe_ptr(sqp, sqp->s_last);
+
+	/* Return if it is not OK to start a new work reqeust. */
+	if (!(ib_ipath_state_ops[sqp->state] & IPATH_PROCESS_NEXT_SEND_OK)) {
+		if (!(ib_ipath_state_ops[sqp->state] & IPATH_FLUSH_SEND))
+			goto clr_busy;
+		/* We are in the error state, flush the work request. */
+		send_status = IB_WC_WR_FLUSH_ERR;
+		goto flush_send;
 	}
 
 	/*
 	 * We can rely on the entry not changing without the s_lock
 	 * being held until we update s_last.
+	 * We increment s_cur to indicate s_last is in progress.
 	 */
-	wqe = get_swqe_ptr(sqp, sqp->s_last);
+	if (sqp->s_last == sqp->s_cur) {
+		if (++sqp->s_cur >= sqp->s_size)
+			sqp->s_cur = 0;
+	}
 	spin_unlock_irqrestore(&sqp->s_lock, flags);
 
-	wc.wc_flags = 0;
-	wc.imm_data = 0;
+	if (!qp || !(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_RECV_OK)) {
+		dev->n_pkt_drops++;
+		/*
+		 * For RC, the requester would timeout and retry so
+		 * shortcut the timeouts and just signal too many retries.
+		 */
+		if (sqp->ibqp.qp_type == IB_QPT_RC)
+			send_status = IB_WC_RETRY_EXC_ERR;
+		else
+			send_status = IB_WC_SUCCESS;
+		goto serr;
+	}
+
+	memset(&wc, 0, sizeof wc);
+	send_status = IB_WC_SUCCESS;
 
 	sqp->s_sge.sge = wqe->sg_list[0];
 	sqp->s_sge.sg_list = wqe->sg_list + 1;
@@ -295,69 +332,41 @@ again:
 	switch (wqe->wr.opcode) {
 	case IB_WR_SEND_WITH_IMM:
 		wc.wc_flags = IB_WC_WITH_IMM;
-		wc.imm_data = wqe->wr.imm_data;
+		wc.ex.imm_data = wqe->wr.ex.imm_data;
 		/* FALLTHROUGH */
 	case IB_WR_SEND:
-		if (!ipath_get_rwqe(qp, 0)) {
-		rnr_nak:
-			/* Handle RNR NAK */
-			if (qp->ibqp.qp_type == IB_QPT_UC)
-				goto send_comp;
-			if (sqp->s_rnr_retry == 0) {
-				wc.status = IB_WC_RNR_RETRY_EXC_ERR;
-				goto err;
-			}
-			if (sqp->s_rnr_retry_cnt < 7)
-				sqp->s_rnr_retry--;
-			dev->n_rnr_naks++;
-			sqp->s_rnr_timeout =
-				ib_ipath_rnr_table[sqp->r_min_rnr_timer];
-			ipath_insert_rnr_queue(sqp);
-			goto done;
-		}
+		if (!ipath_get_rwqe(qp, 0))
+			goto rnr_nak;
 		break;
 
 	case IB_WR_RDMA_WRITE_WITH_IMM:
+		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
+			goto inv_err;
 		wc.wc_flags = IB_WC_WITH_IMM;
-		wc.imm_data = wqe->wr.imm_data;
+		wc.ex.imm_data = wqe->wr.ex.imm_data;
 		if (!ipath_get_rwqe(qp, 1))
 			goto rnr_nak;
 		/* FALLTHROUGH */
 	case IB_WR_RDMA_WRITE:
+		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
+			goto inv_err;
 		if (wqe->length == 0)
 			break;
-		if (unlikely(!ipath_rkey_ok(dev, &qp->r_sge, wqe->length,
+		if (unlikely(!ipath_rkey_ok(qp, &qp->r_sge, wqe->length,
 					    wqe->wr.wr.rdma.remote_addr,
 					    wqe->wr.wr.rdma.rkey,
-					    IB_ACCESS_REMOTE_WRITE))) {
-		acc_err:
-			wc.status = IB_WC_REM_ACCESS_ERR;
-		err:
-			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 = sqp->ibqp.qp_num;
-			wc.src_qp = sqp->remote_qpn;
-			wc.pkey_index = 0;
-			wc.slid = sqp->remote_ah_attr.dlid;
-			wc.sl = sqp->remote_ah_attr.sl;
-			wc.dlid_path_bits = 0;
-			wc.port_num = 0;
-			ipath_sqerror_qp(sqp, &wc);
-			goto done;
-		}
+					    IB_ACCESS_REMOTE_WRITE)))
+			goto acc_err;
 		break;
 
 	case IB_WR_RDMA_READ:
-		if (unlikely(!ipath_rkey_ok(dev, &sqp->s_sge, wqe->length,
+		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ)))
+			goto inv_err;
+		if (unlikely(!ipath_rkey_ok(qp, &sqp->s_sge, wqe->length,
 					    wqe->wr.wr.rdma.remote_addr,
 					    wqe->wr.wr.rdma.rkey,
 					    IB_ACCESS_REMOTE_READ)))
 			goto acc_err;
-		if (unlikely(!(qp->qp_access_flags &
-			       IB_ACCESS_REMOTE_READ)))
-			goto acc_err;
 		qp->r_sge.sge = wqe->sg_list[0];
 		qp->r_sge.sg_list = wqe->sg_list + 1;
 		qp->r_sge.num_sge = wqe->wr.num_sge;
@@ -365,26 +374,26 @@ again:
 
 	case IB_WR_ATOMIC_CMP_AND_SWP:
 	case IB_WR_ATOMIC_FETCH_AND_ADD:
-		if (unlikely(!ipath_rkey_ok(dev, &qp->r_sge, sizeof(u64),
-					    wqe->wr.wr.rdma.remote_addr,
-					    wqe->wr.wr.rdma.rkey,
+		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC)))
+			goto inv_err;
+		if (unlikely(!ipath_rkey_ok(qp, &qp->r_sge, sizeof(u64),
+					    wqe->wr.wr.atomic.remote_addr,
+					    wqe->wr.wr.atomic.rkey,
 					    IB_ACCESS_REMOTE_ATOMIC)))
 			goto acc_err;
 		/* Perform atomic OP and save result. */
-		sdata = wqe->wr.wr.atomic.swap;
-		spin_lock_irqsave(&dev->pending_lock, flags);
-		qp->r_atomic_data = *(u64 *) qp->r_sge.sge.vaddr;
-		if (wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)
-			*(u64 *) qp->r_sge.sge.vaddr =
-				qp->r_atomic_data + sdata;
-		else if (qp->r_atomic_data == wqe->wr.wr.atomic.compare_add)
-			*(u64 *) qp->r_sge.sge.vaddr = sdata;
-		spin_unlock_irqrestore(&dev->pending_lock, flags);
-		*(u64 *) sqp->s_sge.sge.vaddr = qp->r_atomic_data;
+		maddr = (atomic64_t *) qp->r_sge.sge.vaddr;
+		sdata = wqe->wr.wr.atomic.compare_add;
+		*(u64 *) sqp->s_sge.sge.vaddr =
+			(wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) ?
+			(u64) atomic64_add_return(sdata, maddr) - sdata :
+			(u64) cmpxchg((u64 *) qp->r_sge.sge.vaddr,
+				      sdata, wqe->wr.wr.atomic.swap);
 		goto send_comp;
 
 	default:
-		goto done;
+		send_status = IB_WC_LOC_QP_OP_ERR;
+		goto serr;
 	}
 
 	sge = &sqp->s_sge.sge;
@@ -393,6 +402,8 @@ again:
 
 		if (len > sge->length)
 			len = sge->length;
+		if (len > sge->sge_length)
+			len = sge->sge_length;
 		BUG_ON(len == 0);
 		ipath_copy_sge(&qp->r_sge, sge->vaddr, len);
 		sge->vaddr += len;
@@ -415,8 +426,7 @@ again:
 		sqp->s_len -= len;
 	}
 
-	if (wqe->wr.opcode == IB_WR_RDMA_WRITE ||
-	    wqe->wr.opcode == IB_WR_RDMA_READ)
+	if (!test_and_clear_bit(IPATH_R_WRID_VALID, &qp->r_aflags))
 		goto send_comp;
 
 	if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM)
@@ -425,58 +435,98 @@ again:
 		wc.opcode = IB_WC_RECV;
 	wc.wr_id = qp->r_wr_id;
 	wc.status = IB_WC_SUCCESS;
-	wc.vendor_err = 0;
 	wc.byte_len = wqe->length;
-	wc.qp_num = qp->ibqp.qp_num;
+	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:
+	spin_lock_irqsave(&sqp->s_lock, flags);
+flush_send:
 	sqp->s_rnr_retry = sqp->s_rnr_retry_cnt;
+	ipath_send_complete(sqp, wqe, send_status);
+	goto again;
 
-	if (!test_bit(IPATH_S_SIGNAL_REQ_WR, &sqp->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 = sqp->ibqp.qp_num;
-		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);
+rnr_nak:
+	/* Handle RNR NAK */
+	if (qp->ibqp.qp_type == IB_QPT_UC)
+		goto send_comp;
+	/*
+	 * Note: we don't need the s_lock held since the BUSY flag
+	 * makes this single threaded.
+	 */
+	if (sqp->s_rnr_retry == 0) {
+		send_status = IB_WC_RNR_RETRY_EXC_ERR;
+		goto serr;
 	}
+	if (sqp->s_rnr_retry_cnt < 7)
+		sqp->s_rnr_retry--;
+	spin_lock_irqsave(&sqp->s_lock, flags);
+	if (!(ib_ipath_state_ops[sqp->state] & IPATH_PROCESS_RECV_OK))
+		goto clr_busy;
+	sqp->s_flags |= IPATH_S_WAITING;
+	dev->n_rnr_naks++;
+	sqp->s_rnr_timeout = ib_ipath_rnr_table[qp->r_min_rnr_timer];
+	ipath_insert_rnr_queue(sqp);
+	goto clr_busy;
+
+inv_err:
+	send_status = IB_WC_REM_INV_REQ_ERR;
+	wc.status = IB_WC_LOC_QP_OP_ERR;
+	goto err;
+
+acc_err:
+	send_status = IB_WC_REM_ACCESS_ERR;
+	wc.status = IB_WC_LOC_PROT_ERR;
+err:
+	/* responder goes to error state */
+	ipath_rc_error(qp, wc.status);
 
-	/* Update s_last now that we are finished with the SWQE */
+serr:
 	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);
-	goto again;
+	ipath_send_complete(sqp, wqe, send_status);
+	if (sqp->ibqp.qp_type == IB_QPT_RC) {
+		int lastwqe = ipath_error_qp(sqp, IB_WC_WR_FLUSH_ERR);
+
+		sqp->s_flags &= ~IPATH_S_BUSY;
+		spin_unlock_irqrestore(&sqp->s_lock, flags);
+		if (lastwqe) {
+			struct ib_event ev;
 
+			ev.device = sqp->ibqp.device;
+			ev.element.qp = &sqp->ibqp;
+			ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
+			sqp->ibqp.event_handler(&ev, sqp->ibqp.qp_context);
+		}
+		goto done;
+	}
+clr_busy:
+	sqp->s_flags &= ~IPATH_S_BUSY;
+unlock:
+	spin_unlock_irqrestore(&sqp->s_lock, flags);
 done:
-	if (atomic_dec_and_test(&qp->refcount))
+	if (qp && atomic_dec_and_test(&qp->refcount))
 		wake_up(&qp->wait);
 }
 
-static int want_buffer(struct ipath_devdata *dd)
+static void want_buffer(struct ipath_devdata *dd, struct ipath_qp *qp)
 {
-	set_bit(IPATH_S_PIOINTBUFAVAIL, &dd->ipath_sendctrl);
-	ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
-			 dd->ipath_sendctrl);
-
-	return 0;
+	if (!(dd->ipath_flags & IPATH_HAS_SEND_DMA) ||
+	    qp->ibqp.qp_type == IB_QPT_SMI) {
+		unsigned long flags;
+
+		spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
+		dd->ipath_sendctrl |= INFINIPATH_S_PIOINTBUFAVAIL;
+		ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
+				 dd->ipath_sendctrl);
+		ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
+		spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
+	}
 }
 
 /**
@@ -485,115 +535,35 @@ static int want_buffer(struct ipath_devdata *dd)
  * @dev: the device we ran out of buffers on
  *
  * Called when we run out of PIO buffers.
+ * If we are now in the error state, return zero to flush the
+ * send work request.
  */
-void ipath_no_bufs_available(struct ipath_qp *qp, struct ipath_ibdev *dev)
+static int ipath_no_bufs_available(struct ipath_qp *qp,
+				    struct ipath_ibdev *dev)
 {
 	unsigned long flags;
+	int ret = 1;
 
-	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()
-	 * could be called.  If we are still in the tasklet function,
-	 * tasklet_hi_schedule() will not call us until the next time
-	 * tasklet_hi_schedule() is called.
-	 * We clear the tasklet flag now since we are committing to return
-	 * from the tasklet function.
+	 * could be called. Therefore, put QP on the piowait list before
+	 * enabling the PIO avail interrupt.
 	 */
-	clear_bit(IPATH_S_BUSY, &qp->s_flags);
-	tasklet_unlock(&qp->s_task);
-	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;
-	}
-	/* 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(&to_idev(qp->ibqp.device)->lk_table,
-				   &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;
+	if (ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK) {
+		dev->n_piowait++;
+		qp->s_flags |= IPATH_S_WAITING;
+		qp->s_flags &= ~IPATH_S_BUSY;
+		spin_lock(&dev->pending_lock);
+		if (list_empty(&qp->piowait))
+			list_add_tail(&qp->piowait, &dev->piowait);
+		spin_unlock(&dev->pending_lock);
+	} else
+		ret = 0;
 	spin_unlock_irqrestore(&qp->s_lock, flags);
-
-	ipath_do_ruc_send((unsigned long) qp);
-
-	ret = 0;
-
-bail:
+	if (ret)
+		want_buffer(dev->dd, qp);
 	return ret;
 }
 
@@ -627,39 +597,78 @@ 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 |
+				       qp->remote_ah_attr.src_path_bits);
+	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);
+	int (*make_req)(struct ipath_qp *qp);
 	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))
+	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 bail;
+	}
 
-	if (unlikely(qp->remote_ah_attr.dlid == dev->dd->ipath_lid)) {
-		ipath_ruc_loopback(qp);
-		goto clear;
+	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;
+
+	spin_lock_irqsave(&qp->s_lock, flags);
+
+	/* Return if we are already busy processing a work request. */
+	if ((qp->s_flags & (IPATH_S_BUSY | IPATH_S_ANY_WAIT)) ||
+	    !(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_OR_FLUSH_SEND)) {
+		spin_unlock_irqrestore(&qp->s_lock, flags);
+		goto bail;
 	}
 
-	ohdr = &qp->s_hdr.u.oth;
-	if (qp->remote_ah_attr.ah_flags & IB_AH_GRH)
-		ohdr = &qp->s_hdr.u.l.oth;
+	qp->s_flags |= IPATH_S_BUSY;
+
+	spin_unlock_irqrestore(&qp->s_lock, flags);
 
 again:
 	/* Check for a constructed packet to be sent. */
@@ -668,69 +677,58 @@ 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)) {
-			ipath_no_bufs_available(qp, dev);
-			goto bail;
+		if (ipath_verbs_send(qp, &qp->s_hdr, qp->s_hdrwords,
+				     qp->s_cur_sge, qp->s_cur_size)) {
+			if (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);
+	if (make_req(qp))
+		goto again;
 
-	/* Sending responses has higher priority over sending requests. */
-	if (qp->s_ack_state != IB_OPCODE_RC_ACKNOWLEDGE &&
-	    (bth0 = ipath_make_rc_ack(qp, ohdr, pmtu)) != 0)
-		bth2 = qp->s_ack_psn++ & IPATH_PSN_MASK;
-	else 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_flags);
-		spin_unlock_irqrestore(&qp->s_lock, flags);
-		goto bail;
-	}
+bail:;
+}
 
-	spin_unlock_irqrestore(&qp->s_lock, flags);
+/*
+ * This should be called with s_lock held.
+ */
+void ipath_send_complete(struct ipath_qp *qp, struct ipath_swqe *wqe,
+			 enum ib_wc_status status)
+{
+	u32 old_last, 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 (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_OR_FLUSH_SEND))
+		return;
 
-	/* 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_flags);
-bail:
-	return;
+		memset(&wc, 0, sizeof wc);
+		wc.wr_id = wqe->wr.wr_id;
+		wc.status = status;
+		wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
+		wc.qp = &qp->ibqp;
+		if (status == IB_WC_SUCCESS)
+			wc.byte_len = wqe->length;
+		ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc,
+			       status != IB_WC_SUCCESS);
+	}
+
+	old_last = last = qp->s_last;
+	if (++last >= qp->s_size)
+		last = 0;
+	qp->s_last = last;
+	if (qp->s_cur == old_last)
+		qp->s_cur = last;
+	if (qp->s_tail == old_last)
+		qp->s_tail = last;
+	if (qp->state == IB_QPS_SQD && last == qp->s_cur)
+		qp->s_draining = 0;
 }