IB/ipath: Fix many locking issues when switching to error state

The send DMA hardware queue voided a number of prior assumptions about
when a send is complete which led to completions being generated out of
order.  There were also a number of locking issues when switching the QP
to the error or reset states, and we implement the IB_QPS_SQD state.
Signed-off-by: Ralph Campbell <ralph.campbell@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>

drivers/infiniband/hw/ipath/ipath_qp.c

@@ -242,7 +242,6 @@ static void ipath_free_qp(struct ipath_qp_table *qpt, struct ipath_qp *qp)
 {
 	struct ipath_qp *q, **qpp;
 	unsigned long flags;
-	int fnd = 0;
 
 	spin_lock_irqsave(&qpt->lock, flags);
 
@@ -253,51 +252,40 @@ static void ipath_free_qp(struct ipath_qp_table *qpt, struct ipath_qp *qp)
 			*qpp = qp->next;
 			qp->next = NULL;
 			atomic_dec(&qp->refcount);
-			fnd = 1;
 			break;
 		}
 	}
 
 	spin_unlock_irqrestore(&qpt->lock, flags);
-
-	if (!fnd)
-		return;
-
-	free_qpn(qpt, qp->ibqp.qp_num);
-
-	wait_event(qp->wait, !atomic_read(&qp->refcount));
 }
 
 /**
- * ipath_free_all_qps - remove all QPs from the table
+ * ipath_free_all_qps - check for QPs still in use
  * @qpt: the QP table to empty
+ *
+ * There should not be any QPs still in use.
+ * Free memory for table.
  */
-void ipath_free_all_qps(struct ipath_qp_table *qpt)
+unsigned ipath_free_all_qps(struct ipath_qp_table *qpt)
 {
 	unsigned long flags;
-	struct ipath_qp *qp, *nqp;
-	u32 n;
+	struct ipath_qp *qp;
+	u32 n, qp_inuse = 0;
 
+	spin_lock_irqsave(&qpt->lock, flags);
 	for (n = 0; n < qpt->max; n++) {
-		spin_lock_irqsave(&qpt->lock, flags);
 		qp = qpt->table[n];
 		qpt->table[n] = NULL;
-		spin_unlock_irqrestore(&qpt->lock, flags);
-
-		while (qp) {
-			nqp = qp->next;
-			free_qpn(qpt, qp->ibqp.qp_num);
-			if (!atomic_dec_and_test(&qp->refcount) ||
-			    !ipath_destroy_qp(&qp->ibqp))
-				ipath_dbg("QP memory leak!\n");
-			qp = nqp;
-		}
+
+		for (; qp; qp = qp->next)
+			qp_inuse++;
 	}
+	spin_unlock_irqrestore(&qpt->lock, flags);
 
-	for (n = 0; n < ARRAY_SIZE(qpt->map); n++) {
+	for (n = 0; n < ARRAY_SIZE(qpt->map); n++)
 		if (qpt->map[n].page)
-			free_page((unsigned long)qpt->map[n].page);
-	}
+			free_page((unsigned long) qpt->map[n].page);
+	return qp_inuse;
 }
 
 /**
@@ -336,11 +324,12 @@ static void ipath_reset_qp(struct ipath_qp *qp, enum ib_qp_type type)
 	qp->remote_qpn = 0;
 	qp->qkey = 0;
 	qp->qp_access_flags = 0;
-	qp->s_busy = 0;
+	atomic_set(&qp->s_dma_busy, 0);
 	qp->s_flags &= IPATH_S_SIGNAL_REQ_WR;
 	qp->s_hdrwords = 0;
 	qp->s_wqe = NULL;
 	qp->s_pkt_delay = 0;
+	qp->s_draining = 0;
 	qp->s_psn = 0;
 	qp->r_psn = 0;
 	qp->r_msn = 0;
@@ -353,7 +342,8 @@ static void ipath_reset_qp(struct ipath_qp *qp, enum ib_qp_type type)
 	}
 	qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE;
 	qp->r_nak_state = 0;
-	qp->r_wrid_valid = 0;
+	qp->r_aflags = 0;
+	qp->r_flags = 0;
 	qp->s_rnr_timeout = 0;
 	qp->s_head = 0;
 	qp->s_tail = 0;
@@ -361,7 +351,6 @@ static void ipath_reset_qp(struct ipath_qp *qp, enum ib_qp_type type)
 	qp->s_last = 0;
 	qp->s_ssn = 1;
 	qp->s_lsn = 0;
-	qp->s_wait_credit = 0;
 	memset(qp->s_ack_queue, 0, sizeof(qp->s_ack_queue));
 	qp->r_head_ack_queue = 0;
 	qp->s_tail_ack_queue = 0;
@@ -370,7 +359,6 @@ static void ipath_reset_qp(struct ipath_qp *qp, enum ib_qp_type type)
 		qp->r_rq.wq->head = 0;
 		qp->r_rq.wq->tail = 0;
 	}
-	qp->r_reuse_sge = 0;
 }
 
 /**
@@ -402,39 +390,21 @@ int ipath_error_qp(struct ipath_qp *qp, enum ib_wc_status err)
 		list_del_init(&qp->piowait);
 	spin_unlock(&dev->pending_lock);
 
-	wc.vendor_err = 0;
-	wc.byte_len = 0;
-	wc.imm_data = 0;
+	/* Schedule the sending tasklet to drain the send work queue. */
+	if (qp->s_last != qp->s_head)
+		ipath_schedule_send(qp);
+
+	memset(&wc, 0, sizeof(wc));
 	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;
-	if (qp->r_wrid_valid) {
-		qp->r_wrid_valid = 0;
+	wc.opcode = IB_WC_RECV;
+
+	if (test_and_clear_bit(IPATH_R_WRID_VALID, &qp->r_aflags)) {
 		wc.wr_id = qp->r_wr_id;
-		wc.opcode = IB_WC_RECV;
 		wc.status = err;
 		ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
 	}
 	wc.status = IB_WC_WR_FLUSH_ERR;
 
-	while (qp->s_last != qp->s_head) {
-		struct ipath_swqe *wqe = get_swqe_ptr(qp, qp->s_last);
-
-		wc.wr_id = wqe->wr.wr_id;
-		wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
-		if (++qp->s_last >= qp->s_size)
-			qp->s_last = 0;
-		ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 1);
-	}
-	qp->s_cur = qp->s_tail = qp->s_head;
-	qp->s_hdrwords = 0;
-	qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE;
-
 	if (qp->r_rq.wq) {
 		struct ipath_rwq *wq;
 		u32 head;
@@ -450,7 +420,6 @@ int ipath_error_qp(struct ipath_qp *qp, enum ib_wc_status err)
 		tail = wq->tail;
 		if (tail >= qp->r_rq.size)
 			tail = 0;
-		wc.opcode = IB_WC_RECV;
 		while (tail != head) {
 			wc.wr_id = get_rwqe_ptr(&qp->r_rq, tail)->wr_id;
 			if (++tail >= qp->r_rq.size)
@@ -482,11 +451,10 @@ int ipath_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 	struct ipath_ibdev *dev = to_idev(ibqp->device);
 	struct ipath_qp *qp = to_iqp(ibqp);
 	enum ib_qp_state cur_state, new_state;
-	unsigned long flags;
 	int lastwqe = 0;
 	int ret;
 
-	spin_lock_irqsave(&qp->s_lock, flags);
+	spin_lock_irq(&qp->s_lock);
 
 	cur_state = attr_mask & IB_QP_CUR_STATE ?
 		attr->cur_qp_state : qp->state;
@@ -539,16 +507,42 @@ int ipath_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 
 	switch (new_state) {
 	case IB_QPS_RESET:
+		if (qp->state != IB_QPS_RESET) {
+			qp->state = IB_QPS_RESET;
+			spin_lock(&dev->pending_lock);
+			if (!list_empty(&qp->timerwait))
+				list_del_init(&qp->timerwait);
+			if (!list_empty(&qp->piowait))
+				list_del_init(&qp->piowait);
+			spin_unlock(&dev->pending_lock);
+			qp->s_flags &= ~IPATH_S_ANY_WAIT;
+			spin_unlock_irq(&qp->s_lock);
+			/* Stop the sending tasklet */
+			tasklet_kill(&qp->s_task);
+			wait_event(qp->wait_dma, !atomic_read(&qp->s_dma_busy));
+			spin_lock_irq(&qp->s_lock);
+		}
 		ipath_reset_qp(qp, ibqp->qp_type);
 		break;
 
+	case IB_QPS_SQD:
+		qp->s_draining = qp->s_last != qp->s_cur;
+		qp->state = new_state;
+		break;
+
+	case IB_QPS_SQE:
+		if (qp->ibqp.qp_type == IB_QPT_RC)
+			goto inval;
+		qp->state = new_state;
+		break;
+
 	case IB_QPS_ERR:
 		lastwqe = ipath_error_qp(qp, IB_WC_WR_FLUSH_ERR);
 		break;
 
 	default:
+		qp->state = new_state;
 		break;
-
 	}
 
 	if (attr_mask & IB_QP_PKEY_INDEX)
@@ -601,8 +595,7 @@ int ipath_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC)
 		qp->s_max_rd_atomic = attr->max_rd_atomic;
 
-	qp->state = new_state;
-	spin_unlock_irqrestore(&qp->s_lock, flags);
+	spin_unlock_irq(&qp->s_lock);
 
 	if (lastwqe) {
 		struct ib_event ev;
@@ -616,7 +609,7 @@ int ipath_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 	goto bail;
 
 inval:
-	spin_unlock_irqrestore(&qp->s_lock, flags);
+	spin_unlock_irq(&qp->s_lock);
 	ret = -EINVAL;
 
 bail:
@@ -647,7 +640,7 @@ int ipath_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 	attr->pkey_index = qp->s_pkey_index;
 	attr->alt_pkey_index = 0;
 	attr->en_sqd_async_notify = 0;
-	attr->sq_draining = 0;
+	attr->sq_draining = qp->s_draining;
 	attr->max_rd_atomic = qp->s_max_rd_atomic;
 	attr->max_dest_rd_atomic = qp->r_max_rd_atomic;
 	attr->min_rnr_timer = qp->r_min_rnr_timer;
@@ -837,6 +830,7 @@ 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);
+		init_waitqueue_head(&qp->wait_dma);
 		tasklet_init(&qp->s_task, ipath_do_send, (unsigned long)qp);
 		INIT_LIST_HEAD(&qp->piowait);
 		INIT_LIST_HEAD(&qp->timerwait);
@@ -930,6 +924,7 @@ struct ib_qp *ipath_create_qp(struct ib_pd *ibpd,
 	else
 		vfree(qp->r_rq.wq);
 	ipath_free_qp(&dev->qp_table, qp);
+	free_qpn(&dev->qp_table, qp->ibqp.qp_num);
 bail_qp:
 	kfree(qp);
 bail_swq:
@@ -951,41 +946,44 @@ int ipath_destroy_qp(struct ib_qp *ibqp)
 {
 	struct ipath_qp *qp = to_iqp(ibqp);
 	struct ipath_ibdev *dev = to_idev(ibqp->device);
-	unsigned long flags;
 
-	spin_lock_irqsave(&qp->s_lock, flags);
-	qp->state = IB_QPS_ERR;
-	spin_unlock_irqrestore(&qp->s_lock, flags);
-	spin_lock(&dev->n_qps_lock);
-	dev->n_qps_allocated--;
-	spin_unlock(&dev->n_qps_lock);
+	/* Make sure HW and driver activity is stopped. */
+	spin_lock_irq(&qp->s_lock);
+	if (qp->state != IB_QPS_RESET) {
+		qp->state = IB_QPS_RESET;
+		spin_lock(&dev->pending_lock);
+		if (!list_empty(&qp->timerwait))
+			list_del_init(&qp->timerwait);
+		if (!list_empty(&qp->piowait))
+			list_del_init(&qp->piowait);
+		spin_unlock(&dev->pending_lock);
+		qp->s_flags &= ~IPATH_S_ANY_WAIT;
+		spin_unlock_irq(&qp->s_lock);
+		/* Stop the sending tasklet */
+		tasklet_kill(&qp->s_task);
+		wait_event(qp->wait_dma, !atomic_read(&qp->s_dma_busy));
+	} else
+		spin_unlock_irq(&qp->s_lock);
 
-	/* Stop the sending tasklet. */
-	tasklet_kill(&qp->s_task);
+	ipath_free_qp(&dev->qp_table, qp);
 
 	if (qp->s_tx) {
 		atomic_dec(&qp->refcount);
 		if (qp->s_tx->txreq.flags & IPATH_SDMA_TXREQ_F_FREEBUF)
 			kfree(qp->s_tx->txreq.map_addr);
+		spin_lock_irq(&dev->pending_lock);
+		list_add(&qp->s_tx->txreq.list, &dev->txreq_free);
+		spin_unlock_irq(&dev->pending_lock);
+		qp->s_tx = NULL;
 	}
 
-	/* Make sure the QP isn't on the timeout list. */
-	spin_lock_irqsave(&dev->pending_lock, flags);
-	if (!list_empty(&qp->timerwait))
-		list_del_init(&qp->timerwait);
-	if (!list_empty(&qp->piowait))
-		list_del_init(&qp->piowait);
-	if (qp->s_tx)
-		list_add(&qp->s_tx->txreq.list, &dev->txreq_free);
-	spin_unlock_irqrestore(&dev->pending_lock, flags);
+	wait_event(qp->wait, !atomic_read(&qp->refcount));
 
-	/*
-	 * Make sure that the QP is not in the QPN table so receive
-	 * interrupts will discard packets for this QP.  XXX Also remove QP
-	 * from multicast table.
-	 */
-	if (atomic_read(&qp->refcount) != 0)
-		ipath_free_qp(&dev->qp_table, qp);
+	/* all user's cleaned up, mark it available */
+	free_qpn(&dev->qp_table, qp->ibqp.qp_num);
+	spin_lock(&dev->n_qps_lock);
+	dev->n_qps_allocated--;
+	spin_unlock(&dev->n_qps_lock);
 
 	if (qp->ip)
 		kref_put(&qp->ip->ref, ipath_release_mmap_info);
@@ -1055,9 +1053,10 @@ void ipath_get_credit(struct ipath_qp *qp, u32 aeth)
 	}
 
 	/* Restart sending if it was blocked due to lack of credits. */
-	if (qp->s_cur != qp->s_head &&
+	if ((qp->s_flags & IPATH_S_WAIT_SSN_CREDIT) &&
+	    qp->s_cur != qp->s_head &&
 	    (qp->s_lsn == (u32) -1 ||
 	     ipath_cmp24(get_swqe_ptr(qp, qp->s_cur)->ssn,
 			 qp->s_lsn + 1) <= 0))
-		tasklet_hi_schedule(&qp->s_task);
+		ipath_schedule_send(qp);
 }

drivers/infiniband/hw/ipath/ipath_ruc.c

@@ -78,6 +78,7 @@ 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.
@@ -85,9 +86,9 @@ 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 {
@@ -109,7 +110,7 @@ void ipath_insert_rnr_queue(struct ipath_qp *qp)
 			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);
 }
 
 /**
@@ -185,6 +186,11 @@ 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. */
@@ -192,9 +198,8 @@ 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();
@@ -207,7 +212,7 @@ int ipath_get_rwqe(struct ipath_qp *qp, int wr_id_only)
 	wq->tail = tail;
 
 	ret = 1;
-	qp->r_wrid_valid = 1;
+	set_bit(IPATH_R_WRID_VALID, &qp->r_aflags);
 	if (handler) {
 		u32 n;
 
@@ -234,8 +239,8 @@ int ipath_get_rwqe(struct ipath_qp *qp, int wr_id_only)
 			goto bail;
 		}
 	}
+unlock:
 	spin_unlock_irqrestore(&rq->lock, flags);
-
 bail:
 	return ret;
 }
@@ -263,35 +268,59 @@ static void ipath_ruc_loopback(struct ipath_qp *sqp)
 	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) ||
-	    sqp->s_rnr_timeout) {
-		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);
 
+	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;
 
@@ -396,8 +425,7 @@ static void ipath_ruc_loopback(struct ipath_qp *sqp)
 		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)
@@ -417,6 +445,8 @@ static void ipath_ruc_loopback(struct ipath_qp *sqp)
 		       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;
@@ -437,11 +467,12 @@ static void ipath_ruc_loopback(struct ipath_qp *sqp)
 		sqp->s_rnr_retry--;
 	spin_lock_irqsave(&sqp->s_lock, flags);
 	if (!(ib_ipath_state_ops[sqp->state] & IPATH_PROCESS_RECV_OK))
-		goto unlock;
+		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 unlock;
+	goto clr_busy;
 
 inv_err:
 	send_status = IB_WC_REM_INV_REQ_ERR;
@@ -473,17 +504,19 @@ static void ipath_ruc_loopback(struct ipath_qp *sqp)
 		}
 		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 void want_buffer(struct ipath_devdata *dd, struct ipath_qp *qp)
 {
 	if (!(dd->ipath_flags & IPATH_HAS_SEND_DMA) ||
-		qp->ibqp.qp_type == IB_QPT_SMI) {
+	    qp->ibqp.qp_type == IB_QPT_SMI) {
 		unsigned long flags;
 
 		spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
@@ -501,26 +534,36 @@ static void want_buffer(struct ipath_devdata *dd, struct ipath_qp *qp)
  * @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.
  */
-static void ipath_no_bufs_available(struct ipath_qp *qp,
+static int ipath_no_bufs_available(struct ipath_qp *qp,
 				    struct ipath_ibdev *dev)
 {
 	unsigned long flags;
+	int ret = 1;
 
 	/*
 	 * 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 leave the busy flag set so that another post send doesn't
-	 * try to put the same QP on the piowait list again.
+	 * could be called. Therefore, put QP on the piowait list before
+	 * enabling the PIO avail interrupt.
 	 */
-	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, qp);
-	dev->n_piowait++;
+	spin_lock_irqsave(&qp->s_lock, flags);
+	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);
+	if (ret)
+		want_buffer(dev->dd, qp);
+	return ret;
 }
 
 /**
@@ -596,15 +639,13 @@ 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);
-
-	if (test_and_set_bit(IPATH_S_BUSY, &qp->s_busy))
-		goto bail;
+	unsigned long 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 clear;
+		goto bail;
 	}
 
 	if (qp->ibqp.qp_type == IB_QPT_RC)
@@ -614,6 +655,19 @@ void ipath_do_send(unsigned long data)
 	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;
+	}
+
+	qp->s_flags |= IPATH_S_BUSY;
+
+	spin_unlock_irqrestore(&qp->s_lock, flags);
+
 again:
 	/* Check for a constructed packet to be sent. */
 	if (qp->s_hdrwords != 0) {
@@ -623,8 +677,8 @@ void ipath_do_send(unsigned long data)
 		 */
 		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;
+			if (ipath_no_bufs_available(qp, dev))
+				goto bail;
 		}
 		dev->n_unicast_xmit++;
 		/* Record that we sent the packet and s_hdr is empty. */
@@ -633,16 +687,20 @@ void ipath_do_send(unsigned long data)
 
 	if (make_req(qp))
 		goto again;
-clear:
-	clear_bit(IPATH_S_BUSY, &qp->s_busy);
+
 bail:;
 }
 
+/*
+ * 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)
 {
-	unsigned long flags;
-	u32 last;
+	u32 old_last, last;
+
+	if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_OR_FLUSH_SEND))
+		return;
 
 	/* See ch. 11.2.4.1 and 10.7.3.1 */
 	if (!(qp->s_flags & IPATH_S_SIGNAL_REQ_WR) ||
@@ -661,10 +719,14 @@ void ipath_send_complete(struct ipath_qp *qp, struct ipath_swqe *wqe,
 			       status != IB_WC_SUCCESS);
 	}
 
-	spin_lock_irqsave(&qp->s_lock, flags);
-	last = qp->s_last;
+	old_last = last = qp->s_last;
 	if (++last >= qp->s_size)
 		last = 0;
 	qp->s_last = last;
-	spin_unlock_irqrestore(&qp->s_lock, flags);
+	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;
 }

drivers/infiniband/hw/ipath/ipath_uc.c

 /*
- * Copyright (c) 2006, 2007 QLogic Corporation. 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
@@ -47,14 +47,30 @@ int ipath_make_uc_req(struct ipath_qp *qp)
 {
 	struct ipath_other_headers *ohdr;
 	struct ipath_swqe *wqe;
+	unsigned long flags;
 	u32 hwords;
 	u32 bth0;
 	u32 len;
 	u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
 	int ret = 0;
 
-	if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK))
+	spin_lock_irqsave(&qp->s_lock, flags);
+
+	if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK)) {
+		if (!(ib_ipath_state_ops[qp->state] & IPATH_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 |= IPATH_S_WAIT_DMA;
+			goto bail;
+		}
+		wqe = get_swqe_ptr(qp, qp->s_last);
+		ipath_send_complete(qp, wqe, IB_WC_WR_FLUSH_ERR);
 		goto done;
+	}
 
 	ohdr = &qp->s_hdr.u.oth;
 	if (qp->remote_ah_attr.ah_flags & IB_AH_GRH)
@@ -69,9 +85,12 @@ int ipath_make_uc_req(struct ipath_qp *qp)
 	qp->s_wqe = NULL;
 	switch (qp->s_state) {
 	default:
+		if (!(ib_ipath_state_ops[qp->state] &
+		    IPATH_PROCESS_NEXT_SEND_OK))
+			goto bail;
 		/* Check if send work queue is empty. */
 		if (qp->s_cur == qp->s_head)
-			goto done;
+			goto bail;
 		/*
 		 * Start a new request.
 		 */
@@ -134,7 +153,7 @@ int ipath_make_uc_req(struct ipath_qp *qp)
 			break;
 
 		default:
-			goto done;
+			goto bail;
 		}
 		break;
 
@@ -194,9 +213,14 @@ int ipath_make_uc_req(struct ipath_qp *qp)
 	ipath_make_ruc_header(to_idev(qp->ibqp.device),
 			      qp, ohdr, bth0 | (qp->s_state << 24),
 			      qp->s_next_psn++ & IPATH_PSN_MASK);
+done:
 	ret = 1;
+	goto unlock;
 
-done:
+bail:
+	qp->s_flags &= ~IPATH_S_BUSY;
+unlock:
+	spin_unlock_irqrestore(&qp->s_lock, flags);
 	return ret;
 }
 
@@ -258,8 +282,7 @@ void ipath_uc_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
 	 */
 	opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
 
-	wc.imm_data = 0;
-	wc.wc_flags = 0;
+	memset(&wc, 0, sizeof wc);
 
 	/* Compare the PSN verses the expected PSN. */
 	if (unlikely(ipath_cmp24(psn, qp->r_psn) != 0)) {
@@ -322,8 +345,8 @@ void ipath_uc_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
 	case OP(SEND_ONLY):
 	case OP(SEND_ONLY_WITH_IMMEDIATE):
 	send_first:
-		if (qp->r_reuse_sge) {
-			qp->r_reuse_sge = 0;
+		if (qp->r_flags & IPATH_R_REUSE_SGE) {
+			qp->r_flags &= ~IPATH_R_REUSE_SGE;
 			qp->r_sge = qp->s_rdma_read_sge;
 		} else if (!ipath_get_rwqe(qp, 0)) {
 			dev->n_pkt_drops++;
@@ -340,13 +363,13 @@ void ipath_uc_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
 	case OP(SEND_MIDDLE):
 		/* Check for invalid length PMTU or posted rwqe len. */
 		if (unlikely(tlen != (hdrsize + pmtu + 4))) {
-			qp->r_reuse_sge = 1;
+			qp->r_flags |= IPATH_R_REUSE_SGE;
 			dev->n_pkt_drops++;
 			goto done;
 		}
 		qp->r_rcv_len += pmtu;
 		if (unlikely(qp->r_rcv_len > qp->r_len)) {
-			qp->r_reuse_sge = 1;
+			qp->r_flags |= IPATH_R_REUSE_SGE;
 			dev->n_pkt_drops++;
 			goto done;
 		}
@@ -372,7 +395,7 @@ void ipath_uc_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
 		/* Check for invalid length. */
 		/* XXX LAST len should be >= 1 */
 		if (unlikely(tlen < (hdrsize + pad + 4))) {
-			qp->r_reuse_sge = 1;
+			qp->r_flags |= IPATH_R_REUSE_SGE;
 			dev->n_pkt_drops++;
 			goto done;
 		}
@@ -380,7 +403,7 @@ void ipath_uc_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
 		tlen -= (hdrsize + pad + 4);
 		wc.byte_len = tlen + qp->r_rcv_len;
 		if (unlikely(wc.byte_len > qp->r_len)) {
-			qp->r_reuse_sge = 1;
+			qp->r_flags |= IPATH_R_REUSE_SGE;
 			dev->n_pkt_drops++;
 			goto done;
 		}
@@ -390,14 +413,10 @@ void ipath_uc_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
 		wc.wr_id = qp->r_wr_id;
 		wc.status = IB_WC_SUCCESS;
 		wc.opcode = IB_WC_RECV;
-		wc.vendor_err = 0;
 		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;
 		/* Signal completion event if the solicited bit is set. */
 		ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
 			       (ohdr->bth[0] &
@@ -488,8 +507,8 @@ void ipath_uc_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
 			dev->n_pkt_drops++;
 			goto done;
 		}
-		if (qp->r_reuse_sge)
-			qp->r_reuse_sge = 0;
+		if (qp->r_flags & IPATH_R_REUSE_SGE)
+			qp->r_flags &= ~IPATH_R_REUSE_SGE;
 		else if (!ipath_get_rwqe(qp, 1)) {
 			dev->n_pkt_drops++;
 			goto done;

drivers/infiniband/hw/ipath/ipath_ud.c

@@ -65,9 +65,9 @@ static void ipath_ud_loopback(struct ipath_qp *sqp, struct ipath_swqe *swqe)
 	u32 length;
 
 	qp = ipath_lookup_qpn(&dev->qp_table, swqe->wr.wr.ud.remote_qpn);
-	if (!qp) {
+	if (!qp || !(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_RECV_OK)) {
 		dev->n_pkt_drops++;
-		goto send_comp;
+		goto done;
 	}
 
 	rsge.sg_list = NULL;
@@ -91,14 +91,12 @@ static void ipath_ud_loopback(struct ipath_qp *sqp, struct ipath_swqe *swqe)
 	 * present on the wire.
 	 */
 	length = swqe->length;
+	memset(&wc, 0, sizeof wc);
 	wc.byte_len = length + sizeof(struct ib_grh);
 
 	if (swqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
 		wc.wc_flags = IB_WC_WITH_IMM;
 		wc.imm_data = swqe->wr.ex.imm_data;
-	} else {
-		wc.wc_flags = 0;
-		wc.imm_data = 0;
 	}
 
 	/*
@@ -229,7 +227,6 @@ static void ipath_ud_loopback(struct ipath_qp *sqp, struct ipath_swqe *swqe)
 	}
 	wc.status = IB_WC_SUCCESS;
 	wc.opcode = IB_WC_RECV;
-	wc.vendor_err = 0;
 	wc.qp = &qp->ibqp;
 	wc.src_qp = sqp->ibqp.qp_num;
 	/* XXX do we know which pkey matched? Only needed for GSI. */
@@ -248,8 +245,7 @@ static void ipath_ud_loopback(struct ipath_qp *sqp, struct ipath_swqe *swqe)
 	kfree(rsge.sg_list);
 	if (atomic_dec_and_test(&qp->refcount))
 		wake_up(&qp->wait);
-send_comp:
-	ipath_send_complete(sqp, swqe, IB_WC_SUCCESS);
+done:;
 }
 
 /**
@@ -264,6 +260,7 @@ int ipath_make_ud_req(struct ipath_qp *qp)
 	struct ipath_other_headers *ohdr;
 	struct ib_ah_attr *ah_attr;
 	struct ipath_swqe *wqe;
+	unsigned long flags;
 	u32 nwords;
 	u32 extra_bytes;
 	u32 bth0;
@@ -271,13 +268,30 @@ int ipath_make_ud_req(struct ipath_qp *qp)
 	u16 lid;
 	int ret = 0;
 
-	if (unlikely(!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK)))
-		goto bail;
+	spin_lock_irqsave(&qp->s_lock, flags);
+
+	if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_NEXT_SEND_OK)) {
+		if (!(ib_ipath_state_ops[qp->state] & IPATH_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 |= IPATH_S_WAIT_DMA;
+			goto bail;
+		}
+		wqe = get_swqe_ptr(qp, qp->s_last);
+		ipath_send_complete(qp, wqe, IB_WC_WR_FLUSH_ERR);
+		goto done;
+	}
 
 	if (qp->s_cur == qp->s_head)
 		goto bail;
 
 	wqe = get_swqe_ptr(qp, qp->s_cur);
+	if (++qp->s_cur >= qp->s_size)
+		qp->s_cur = 0;
 
 	/* Construct the header. */
 	ah_attr = &to_iah(wqe->wr.wr.ud.ah)->attr;
@@ -288,10 +302,23 @@ int ipath_make_ud_req(struct ipath_qp *qp)
 			dev->n_unicast_xmit++;
 	} else {
 		dev->n_unicast_xmit++;
-		lid = ah_attr->dlid &
-			~((1 << dev->dd->ipath_lmc) - 1);
+		lid = ah_attr->dlid & ~((1 << dev->dd->ipath_lmc) - 1);
 		if (unlikely(lid == dev->dd->ipath_lid)) {
+			/*
+			 * If DMAs are in progress, we can't generate
+			 * a completion for the loopback packet since
+			 * it would be out of order.
+			 * XXX Instead of waiting, we could queue a
+			 * zero length descriptor so we get a callback.
+			 */
+			if (atomic_read(&qp->s_dma_busy)) {
+				qp->s_flags |= IPATH_S_WAIT_DMA;
+				goto bail;
+			}
+			spin_unlock_irqrestore(&qp->s_lock, flags);
 			ipath_ud_loopback(qp, wqe);
+			spin_lock_irqsave(&qp->s_lock, flags);
+			ipath_send_complete(qp, wqe, IB_WC_SUCCESS);
 			goto done;
 		}
 	}
@@ -368,11 +395,13 @@ int ipath_make_ud_req(struct ipath_qp *qp)
 	ohdr->u.ud.deth[1] = cpu_to_be32(qp->ibqp.qp_num);
 
 done:
-	if (++qp->s_cur >= qp->s_size)
-		qp->s_cur = 0;
 	ret = 1;
+	goto unlock;
 
 bail:
+	qp->s_flags &= ~IPATH_S_BUSY;
+unlock:
+	spin_unlock_irqrestore(&qp->s_lock, flags);
 	return ret;
 }
 
@@ -506,8 +535,8 @@ void ipath_ud_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
 	/*
 	 * Get the next work request entry to find where to put the data.
 	 */
-	if (qp->r_reuse_sge)
-		qp->r_reuse_sge = 0;
+	if (qp->r_flags & IPATH_R_REUSE_SGE)
+		qp->r_flags &= ~IPATH_R_REUSE_SGE;
 	else if (!ipath_get_rwqe(qp, 0)) {
 		/*
 		 * Count VL15 packets dropped due to no receive buffer.
@@ -523,7 +552,7 @@ void ipath_ud_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
 	}
 	/* Silently drop packets which are too big. */
 	if (wc.byte_len > qp->r_len) {
-		qp->r_reuse_sge = 1;
+		qp->r_flags |= IPATH_R_REUSE_SGE;
 		dev->n_pkt_drops++;
 		goto bail;
 	}
@@ -535,7 +564,8 @@ void ipath_ud_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
 		ipath_skip_sge(&qp->r_sge, sizeof(struct ib_grh));
 	ipath_copy_sge(&qp->r_sge, data,
 		       wc.byte_len - sizeof(struct ib_grh));
-	qp->r_wrid_valid = 0;
+	if (!test_and_clear_bit(IPATH_R_WRID_VALID, &qp->r_aflags))
+		goto bail;
 	wc.wr_id = qp->r_wr_id;
 	wc.status = IB_WC_SUCCESS;
 	wc.opcode = IB_WC_RECV;

drivers/infiniband/hw/ipath/ipath_user_sdma.h

@@ -45,8 +45,6 @@ int ipath_user_sdma_writev(struct ipath_devdata *dd,
 int ipath_user_sdma_make_progress(struct ipath_devdata *dd,
 				  struct ipath_user_sdma_queue *pq);
 
-int ipath_user_sdma_pkt_sent(const struct ipath_user_sdma_queue *pq,
-			     u32 counter);
 void ipath_user_sdma_queue_drain(struct ipath_devdata *dd,
 				 struct ipath_user_sdma_queue *pq);
 

drivers/infiniband/hw/ipath/ipath_verbs.h

@@ -74,6 +74,11 @@
 #define IPATH_POST_RECV_OK		0x02
 #define IPATH_PROCESS_RECV_OK		0x04
 #define IPATH_PROCESS_SEND_OK		0x08
+#define IPATH_PROCESS_NEXT_SEND_OK	0x10
+#define IPATH_FLUSH_SEND		0x20
+#define IPATH_FLUSH_RECV		0x40
+#define IPATH_PROCESS_OR_FLUSH_SEND \
+	(IPATH_PROCESS_SEND_OK | IPATH_FLUSH_SEND)
 
 /* IB Performance Manager status values */
 #define IB_PMA_SAMPLE_STATUS_DONE	0x00
@@ -353,12 +358,14 @@ struct ipath_qp {
 	struct ib_qp ibqp;
 	struct ipath_qp *next;		/* link list for QPN hash table */
 	struct ipath_qp *timer_next;	/* link list for ipath_ib_timer() */
+	struct ipath_qp *pio_next;	/* link for ipath_ib_piobufavail() */
 	struct list_head piowait;	/* link for wait PIO buf */
 	struct list_head timerwait;	/* link for waiting for timeouts */
 	struct ib_ah_attr remote_ah_attr;
 	struct ipath_ib_header s_hdr;	/* next packet header to send */
 	atomic_t refcount;
 	wait_queue_head_t wait;
+	wait_queue_head_t wait_dma;
 	struct tasklet_struct s_task;
 	struct ipath_mmap_info *ip;
 	struct ipath_sge_state *s_cur_sge;
@@ -369,7 +376,7 @@ struct ipath_qp {
 	struct ipath_sge_state s_rdma_read_sge;
 	struct ipath_sge_state r_sge;	/* current receive data */
 	spinlock_t s_lock;
-	unsigned long s_busy;
+	atomic_t s_dma_busy;
 	u16 s_pkt_delay;
 	u16 s_hdrwords;		/* size of s_hdr in 32 bit words */
 	u32 s_cur_size;		/* size of send packet in bytes */
@@ -383,6 +390,7 @@ struct ipath_qp {
 	u32 s_rnr_timeout;	/* number of milliseconds for RNR timeout */
 	u32 r_ack_psn;		/* PSN for next ACK or atomic ACK */
 	u64 r_wr_id;		/* ID for current receive WQE */
+	unsigned long r_aflags;
 	u32 r_len;		/* total length of r_sge */
 	u32 r_rcv_len;		/* receive data len processed */
 	u32 r_psn;		/* expected rcv packet sequence number */
@@ -394,8 +402,7 @@ struct ipath_qp {
 	u8 r_state;		/* opcode of last packet received */
 	u8 r_nak_state;		/* non-zero if NAK is pending */
 	u8 r_min_rnr_timer;	/* retry timeout value for RNR NAKs */
-	u8 r_reuse_sge;		/* for UC receive errors */
-	u8 r_wrid_valid;	/* r_wrid set but CQ entry not yet made */
+	u8 r_flags;
 	u8 r_max_rd_atomic;	/* max number of RDMA read/atomic to receive */
 	u8 r_head_ack_queue;	/* index into s_ack_queue[] */
 	u8 qp_access_flags;
@@ -404,13 +411,13 @@ struct ipath_qp {
 	u8 s_rnr_retry_cnt;
 	u8 s_retry;		/* requester retry counter */
 	u8 s_rnr_retry;		/* requester RNR retry counter */
-	u8 s_wait_credit;	/* limit number of unacked packets sent */
 	u8 s_pkey_index;	/* PKEY index to use */
 	u8 s_max_rd_atomic;	/* max number of RDMA read/atomic to send */
 	u8 s_num_rd_atomic;	/* number of RDMA read/atomic pending */
 	u8 s_tail_ack_queue;	/* index into s_ack_queue[] */
 	u8 s_flags;
 	u8 s_dmult;
+	u8 s_draining;
 	u8 timeout;		/* Timeout for this QP */
 	enum ib_mtu path_mtu;
 	u32 remote_qpn;
@@ -428,16 +435,39 @@ struct ipath_qp {
 	struct ipath_sge r_sg_list[0];	/* verified SGEs */
 };
 
-/* Bit definition for s_busy. */
-#define IPATH_S_BUSY		0
+/*
+ * Atomic bit definitions for r_aflags.
+ */
+#define IPATH_R_WRID_VALID	0
+
+/*
+ * Bit definitions for r_flags.
+ */
+#define IPATH_R_REUSE_SGE	0x01
 
 /*
  * Bit definitions for s_flags.
+ *
+ * IPATH_S_FENCE_PENDING - waiting for all prior RDMA read or atomic SWQEs
+ *			   before processing the next SWQE
+ * IPATH_S_RDMAR_PENDING - waiting for any RDMA read or atomic SWQEs
+ *			   before processing the next SWQE
+ * IPATH_S_WAITING - waiting for RNR timeout or send buffer available.
+ * IPATH_S_WAIT_SSN_CREDIT - waiting for RC credits to process next SWQE
+ * IPATH_S_WAIT_DMA - waiting for send DMA queue to drain before generating
+ *		      next send completion entry not via send DMA.
  */
 #define IPATH_S_SIGNAL_REQ_WR	0x01
 #define IPATH_S_FENCE_PENDING	0x02
 #define IPATH_S_RDMAR_PENDING	0x04
 #define IPATH_S_ACK_PENDING	0x08
+#define IPATH_S_BUSY		0x10
+#define IPATH_S_WAITING		0x20
+#define IPATH_S_WAIT_SSN_CREDIT	0x40
+#define IPATH_S_WAIT_DMA	0x80
+
+#define IPATH_S_ANY_WAIT (IPATH_S_FENCE_PENDING | IPATH_S_RDMAR_PENDING | \
+	IPATH_S_WAITING | IPATH_S_WAIT_SSN_CREDIT | IPATH_S_WAIT_DMA)
 
 #define IPATH_PSN_CREDIT	512
 
@@ -573,13 +603,11 @@ struct ipath_ibdev {
 	u32 n_rnr_naks;
 	u32 n_other_naks;
 	u32 n_timeouts;
-	u32 n_rc_stalls;
 	u32 n_pkt_drops;
 	u32 n_vl15_dropped;
 	u32 n_wqe_errs;
 	u32 n_rdma_dup_busy;
 	u32 n_piowait;
-	u32 n_no_piobuf;
 	u32 n_unaligned;
 	u32 port_cap_flags;
 	u32 pma_sample_start;
@@ -657,6 +685,17 @@ static inline struct ipath_ibdev *to_idev(struct ib_device *ibdev)
 	return container_of(ibdev, struct ipath_ibdev, ibdev);
 }
 
+/*
+ * This must be called with s_lock held.
+ */
+static inline void ipath_schedule_send(struct ipath_qp *qp)
+{
+	if (qp->s_flags & IPATH_S_ANY_WAIT)
+		qp->s_flags &= ~IPATH_S_ANY_WAIT;
+	if (!(qp->s_flags & IPATH_S_BUSY))
+		tasklet_hi_schedule(&qp->s_task);
+}
+
 int ipath_process_mad(struct ib_device *ibdev,
 		      int mad_flags,
 		      u8 port_num,
@@ -706,7 +745,7 @@ int ipath_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 int ipath_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 		   int attr_mask, struct ib_qp_init_attr *init_attr);
 
-void ipath_free_all_qps(struct ipath_qp_table *qpt);
+unsigned ipath_free_all_qps(struct ipath_qp_table *qpt);
 
 int ipath_init_qp_table(struct ipath_ibdev *idev, int size);