kernel_samsung_a34x-permissive/drivers/infiniband/hw/qib/qib_ruc.c

/*
 * Copyright (c) 2006, 2007, 2008, 2009 QLogic Corporation. All rights reserved.
 * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
 *
 * This software is available to you under a choice of one of two
 * licenses.  You may choose to be licensed under the terms of the GNU
 * General Public License (GPL) Version 2, available from the file
 * COPYING in the main directory of this source tree, or the
 * OpenIB.org BSD license below:
 *
 *     Redistribution and use in source and binary forms, with or
 *     without modification, are permitted provided that the following
 *     conditions are met:
 *
 *      - Redistributions of source code must retain the above
 *        copyright notice, this list of conditions and the following
 *        disclaimer.
 *
 *      - Redistributions in binary form must reproduce the above
 *        copyright notice, this list of conditions and the following
 *        disclaimer in the documentation and/or other materials
 *        provided with the distribution.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 * SOFTWARE.
 */

#include <linux/spinlock.h>
#include <rdma/ib_smi.h>

#include "qib.h"
#include "qib_mad.h"

/*
 * Switch to alternate path.
 * The QP s_lock should be held and interrupts disabled.
 */
void qib_migrate_qp(struct rvt_qp *qp)
{
	struct ib_event ev;

	qp->s_mig_state = IB_MIG_MIGRATED;
	qp->remote_ah_attr = qp->alt_ah_attr;
	qp->port_num = rdma_ah_get_port_num(&qp->alt_ah_attr);
	qp->s_pkey_index = qp->s_alt_pkey_index;

	ev.device = qp->ibqp.device;
	ev.element.qp = &qp->ibqp;
	ev.event = IB_EVENT_PATH_MIG;
	qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
}

static __be64 get_sguid(struct qib_ibport *ibp, unsigned index)
{
	if (!index) {
		struct qib_pportdata *ppd = ppd_from_ibp(ibp);

		return ppd->guid;
	}
	return ibp->guids[index - 1];
}

static int gid_ok(union ib_gid *gid, __be64 gid_prefix, __be64 id)
{
	return (gid->global.interface_id == id &&
		(gid->global.subnet_prefix == gid_prefix ||
		 gid->global.subnet_prefix == IB_DEFAULT_GID_PREFIX));
}

/*
 *
 * This should be called with the QP r_lock held.
 *
 * The s_lock will be acquired around the qib_migrate_qp() call.
 */
int qib_ruc_check_hdr(struct qib_ibport *ibp, struct ib_header *hdr,
		      int has_grh, struct rvt_qp *qp, u32 bth0)
{
	__be64 guid;
	unsigned long flags;

	if (qp->s_mig_state == IB_MIG_ARMED && (bth0 & IB_BTH_MIG_REQ)) {
		if (!has_grh) {
			if (rdma_ah_get_ah_flags(&qp->alt_ah_attr) &
			    IB_AH_GRH)
				goto err;
		} else {
			const struct ib_global_route *grh;

			if (!(rdma_ah_get_ah_flags(&qp->alt_ah_attr) &
			      IB_AH_GRH))
				goto err;
			grh = rdma_ah_read_grh(&qp->alt_ah_attr);
			guid = get_sguid(ibp, grh->sgid_index);
			if (!gid_ok(&hdr->u.l.grh.dgid,
				    ibp->rvp.gid_prefix, guid))
				goto err;
			if (!gid_ok(&hdr->u.l.grh.sgid,
			    grh->dgid.global.subnet_prefix,
			    grh->dgid.global.interface_id))
				goto err;
		}
		if (!qib_pkey_ok((u16)bth0,
				 qib_get_pkey(ibp, qp->s_alt_pkey_index))) {
			qib_bad_pkey(ibp,
				     (u16)bth0,
				     (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF,
				     0, qp->ibqp.qp_num,
				     hdr->lrh[3], hdr->lrh[1]);
			goto err;
		}
		/* Validate the SLID. See Ch. 9.6.1.5 and 17.2.8 */
		if ((be16_to_cpu(hdr->lrh[3]) !=
		     rdma_ah_get_dlid(&qp->alt_ah_attr)) ||
		    ppd_from_ibp(ibp)->port !=
			    rdma_ah_get_port_num(&qp->alt_ah_attr))
			goto err;
		spin_lock_irqsave(&qp->s_lock, flags);
		qib_migrate_qp(qp);
		spin_unlock_irqrestore(&qp->s_lock, flags);
	} else {
		if (!has_grh) {
			if (rdma_ah_get_ah_flags(&qp->remote_ah_attr) &
			    IB_AH_GRH)
				goto err;
		} else {
			const struct ib_global_route *grh;

			if (!(rdma_ah_get_ah_flags(&qp->remote_ah_attr) &
			      IB_AH_GRH))
				goto err;
			grh = rdma_ah_read_grh(&qp->remote_ah_attr);
			guid = get_sguid(ibp, grh->sgid_index);
			if (!gid_ok(&hdr->u.l.grh.dgid,
				    ibp->rvp.gid_prefix, guid))
				goto err;
			if (!gid_ok(&hdr->u.l.grh.sgid,
			    grh->dgid.global.subnet_prefix,
			    grh->dgid.global.interface_id))
				goto err;
		}
		if (!qib_pkey_ok((u16)bth0,
				 qib_get_pkey(ibp, qp->s_pkey_index))) {
			qib_bad_pkey(ibp,
				     (u16)bth0,
				     (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF,
				     0, qp->ibqp.qp_num,
				     hdr->lrh[3], hdr->lrh[1]);
			goto err;
		}
		/* Validate the SLID. See Ch. 9.6.1.5 */
		if (be16_to_cpu(hdr->lrh[3]) !=
		    rdma_ah_get_dlid(&qp->remote_ah_attr) ||
		    ppd_from_ibp(ibp)->port != qp->port_num)
			goto err;
		if (qp->s_mig_state == IB_MIG_REARM &&
		    !(bth0 & IB_BTH_MIG_REQ))
			qp->s_mig_state = IB_MIG_ARMED;
	}

	return 0;

err:
	return 1;
}

/**
 * qib_ruc_loopback - handle UC and RC lookback requests
 * @sqp: the sending QP
 *
 * This is called from qib_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
 * receive interrupts since this is a connected protocol and all packets
 * will pass through here.
 */
static void qib_ruc_loopback(struct rvt_qp *sqp)
{
	struct qib_ibport *ibp = to_iport(sqp->ibqp.device, sqp->port_num);
	struct qib_pportdata *ppd = ppd_from_ibp(ibp);
	struct qib_devdata *dd = ppd->dd;
	struct rvt_dev_info *rdi = &dd->verbs_dev.rdi;
	struct rvt_qp *qp;
	struct rvt_swqe *wqe;
	struct rvt_sge *sge;
	unsigned long flags;
	struct ib_wc wc;
	u64 sdata;
	atomic64_t *maddr;
	enum ib_wc_status send_status;
	int release;
	int ret;

	rcu_read_lock();
	/*
	 * Note that we check the responder QP state after
	 * checking the requester's state.
	 */
	qp = rvt_lookup_qpn(rdi, &ibp->rvp, sqp->remote_qpn);
	if (!qp)
		goto done;

	spin_lock_irqsave(&sqp->s_lock, flags);

	/* Return if we are already busy processing a work request. */
	if ((sqp->s_flags & (RVT_S_BUSY | RVT_S_ANY_WAIT)) ||
	    !(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_OR_FLUSH_SEND))
		goto unlock;

	sqp->s_flags |= RVT_S_BUSY;

again:
	if (sqp->s_last == READ_ONCE(sqp->s_head))
		goto clr_busy;
	wqe = rvt_get_swqe_ptr(sqp, sqp->s_last);

	/* Return if it is not OK to start a new work reqeust. */
	if (!(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_NEXT_SEND_OK)) {
		if (!(ib_rvt_state_ops[sqp->state] & RVT_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.
	 */
	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_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) ||
	    qp->ibqp.qp_type != sqp->ibqp.qp_type) {
		ibp->rvp.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;

	release = 1;
	sqp->s_sge.sge = wqe->sg_list[0];
	sqp->s_sge.sg_list = wqe->sg_list + 1;
	sqp->s_sge.num_sge = wqe->wr.num_sge;
	sqp->s_len = wqe->length;
	switch (wqe->wr.opcode) {
	case IB_WR_SEND_WITH_IMM:
		wc.wc_flags = IB_WC_WITH_IMM;
		wc.ex.imm_data = wqe->wr.ex.imm_data;
		/* FALLTHROUGH */
	case IB_WR_SEND:
		ret = rvt_get_rwqe(qp, false);
		if (ret < 0)
			goto op_err;
		if (!ret)
			goto rnr_nak;
		if (wqe->length > qp->r_len)
			goto inv_err;
		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.ex.imm_data = wqe->wr.ex.imm_data;
		ret = rvt_get_rwqe(qp, true);
		if (ret < 0)
			goto op_err;
		if (!ret)
			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(!rvt_rkey_ok(qp, &qp->r_sge.sge, wqe->length,
					  wqe->rdma_wr.remote_addr,
					  wqe->rdma_wr.rkey,
					  IB_ACCESS_REMOTE_WRITE)))
			goto acc_err;
		qp->r_sge.sg_list = NULL;
		qp->r_sge.num_sge = 1;
		qp->r_sge.total_len = wqe->length;
		break;

	case IB_WR_RDMA_READ:
		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ)))
			goto inv_err;
		if (unlikely(!rvt_rkey_ok(qp, &sqp->s_sge.sge, wqe->length,
					  wqe->rdma_wr.remote_addr,
					  wqe->rdma_wr.rkey,
					  IB_ACCESS_REMOTE_READ)))
			goto acc_err;
		release = 0;
		sqp->s_sge.sg_list = NULL;
		sqp->s_sge.num_sge = 1;
		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;
		qp->r_sge.total_len = wqe->length;
		break;

	case IB_WR_ATOMIC_CMP_AND_SWP:
	case IB_WR_ATOMIC_FETCH_AND_ADD:
		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC)))
			goto inv_err;
		if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, sizeof(u64),
					  wqe->atomic_wr.remote_addr,
					  wqe->atomic_wr.rkey,
					  IB_ACCESS_REMOTE_ATOMIC)))
			goto acc_err;
		/* Perform atomic OP and save result. */
		maddr = (atomic64_t *) qp->r_sge.sge.vaddr;
		sdata = wqe->atomic_wr.compare_add;
		*(u64 *) sqp->s_sge.sge.vaddr =
			(wqe->atomic_wr.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->atomic_wr.swap);
		rvt_put_mr(qp->r_sge.sge.mr);
		qp->r_sge.num_sge = 0;
		goto send_comp;

	default:
		send_status = IB_WC_LOC_QP_OP_ERR;
		goto serr;
	}

	sge = &sqp->s_sge.sge;
	while (sqp->s_len) {
		u32 len = sqp->s_len;

		if (len > sge->length)
			len = sge->length;
		if (len > sge->sge_length)
			len = sge->sge_length;
		BUG_ON(len == 0);
		qib_copy_sge(&qp->r_sge, sge->vaddr, len, release);
		sge->vaddr += len;
		sge->length -= len;
		sge->sge_length -= len;
		if (sge->sge_length == 0) {
			if (!release)
				rvt_put_mr(sge->mr);
			if (--sqp->s_sge.num_sge)
				*sge = *sqp->s_sge.sg_list++;
		} else if (sge->length == 0 && sge->mr->lkey) {
			if (++sge->n >= RVT_SEGSZ) {
				if (++sge->m >= sge->mr->mapsz)
					break;
				sge->n = 0;
			}
			sge->vaddr =
				sge->mr->map[sge->m]->segs[sge->n].vaddr;
			sge->length =
				sge->mr->map[sge->m]->segs[sge->n].length;
		}
		sqp->s_len -= len;
	}
	if (release)
		rvt_put_ss(&qp->r_sge);

	if (!test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags))
		goto send_comp;

	if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM)
		wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
	else
		wc.opcode = IB_WC_RECV;
	wc.wr_id = qp->r_wr_id;
	wc.status = IB_WC_SUCCESS;
	wc.byte_len = wqe->length;
	wc.qp = &qp->ibqp;
	wc.src_qp = qp->remote_qpn;
	wc.slid = rdma_ah_get_dlid(&qp->remote_ah_attr);
	wc.sl = rdma_ah_get_sl(&qp->remote_ah_attr);
	wc.port_num = 1;
	/* Signal completion event if the solicited bit is set. */
	rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc,
		     wqe->wr.send_flags & IB_SEND_SOLICITED);

send_comp:
	spin_lock_irqsave(&sqp->s_lock, flags);
	ibp->rvp.n_loop_pkts++;
flush_send:
	sqp->s_rnr_retry = sqp->s_rnr_retry_cnt;
	qib_send_complete(sqp, wqe, send_status);
	goto again;

rnr_nak:
	/* Handle RNR NAK */
	if (qp->ibqp.qp_type == IB_QPT_UC)
		goto send_comp;
	ibp->rvp.n_rnr_naks++;
	/*
	 * 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_rvt_state_ops[sqp->state] & RVT_PROCESS_RECV_OK))
		goto clr_busy;
	rvt_add_rnr_timer(sqp, qp->r_min_rnr_timer <<
				IB_AETH_CREDIT_SHIFT);
	goto clr_busy;

op_err:
	send_status = IB_WC_REM_OP_ERR;
	wc.status = IB_WC_LOC_QP_OP_ERR;
	goto err;

inv_err:
	send_status =
		sqp->ibqp.qp_type == IB_QPT_RC ?
			IB_WC_REM_INV_REQ_ERR :
			IB_WC_SUCCESS;
	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 */
	rvt_rc_error(qp, wc.status);

serr:
	spin_lock_irqsave(&sqp->s_lock, flags);
	qib_send_complete(sqp, wqe, send_status);
	if (sqp->ibqp.qp_type == IB_QPT_RC) {
		int lastwqe = rvt_error_qp(sqp, IB_WC_WR_FLUSH_ERR);

		sqp->s_flags &= ~RVT_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 &= ~RVT_S_BUSY;
unlock:
	spin_unlock_irqrestore(&sqp->s_lock, flags);
done:
	rcu_read_unlock();
}

/**
 * qib_make_grh - construct a GRH header
 * @ibp: a pointer to the IB port
 * @hdr: a pointer to the GRH header being constructed
 * @grh: the global route address to send to
 * @hwords: the number of 32 bit words of header being sent
 * @nwords: the number of 32 bit words of data being sent
 *
 * Return the size of the header in 32 bit words.
 */
u32 qib_make_grh(struct qib_ibport *ibp, struct ib_grh *hdr,
		 const struct ib_global_route *grh, u32 hwords, u32 nwords)
{
	hdr->version_tclass_flow =
		cpu_to_be32((IB_GRH_VERSION << IB_GRH_VERSION_SHIFT) |
			    (grh->traffic_class << IB_GRH_TCLASS_SHIFT) |
			    (grh->flow_label << IB_GRH_FLOW_SHIFT));
	hdr->paylen = cpu_to_be16((hwords - 2 + nwords + SIZE_OF_CRC) << 2);
	/* next_hdr is defined by C8-7 in ch. 8.4.1 */
	hdr->next_hdr = IB_GRH_NEXT_HDR;
	hdr->hop_limit = grh->hop_limit;
	/* The SGID is 32-bit aligned. */
	hdr->sgid.global.subnet_prefix = ibp->rvp.gid_prefix;
	if (!grh->sgid_index)
		hdr->sgid.global.interface_id = ppd_from_ibp(ibp)->guid;
	else if (grh->sgid_index < QIB_GUIDS_PER_PORT)
		hdr->sgid.global.interface_id = ibp->guids[grh->sgid_index - 1];
	hdr->dgid = grh->dgid;

	/* GRH header size in 32-bit words. */
	return sizeof(struct ib_grh) / sizeof(u32);
}

void qib_make_ruc_header(struct rvt_qp *qp, struct ib_other_headers *ohdr,
			 u32 bth0, u32 bth2)
{
	struct qib_qp_priv *priv = qp->priv;
	struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
	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 = QIB_LRH_BTH;
	if (unlikely(rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH)) {
		qp->s_hdrwords +=
			qib_make_grh(ibp, &priv->s_hdr->u.l.grh,
				     rdma_ah_read_grh(&qp->remote_ah_attr),
				     qp->s_hdrwords, nwords);
		lrh0 = QIB_LRH_GRH;
	}
	lrh0 |= ibp->sl_to_vl[rdma_ah_get_sl(&qp->remote_ah_attr)] << 12 |
		rdma_ah_get_sl(&qp->remote_ah_attr) << 4;
	priv->s_hdr->lrh[0] = cpu_to_be16(lrh0);
	priv->s_hdr->lrh[1] =
			cpu_to_be16(rdma_ah_get_dlid(&qp->remote_ah_attr));
	priv->s_hdr->lrh[2] =
			cpu_to_be16(qp->s_hdrwords + nwords + SIZE_OF_CRC);
	priv->s_hdr->lrh[3] =
		cpu_to_be16(ppd_from_ibp(ibp)->lid |
			    rdma_ah_get_path_bits(&qp->remote_ah_attr));
	bth0 |= qib_get_pkey(ibp, qp->s_pkey_index);
	bth0 |= extra_bytes << 20;
	if (qp->s_mig_state == IB_MIG_MIGRATED)
		bth0 |= IB_BTH_MIG_REQ;
	ohdr->bth[0] = cpu_to_be32(bth0);
	ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
	ohdr->bth[2] = cpu_to_be32(bth2);
	this_cpu_inc(ibp->pmastats->n_unicast_xmit);
}

void _qib_do_send(struct work_struct *work)
{
	struct qib_qp_priv *priv = container_of(work, struct qib_qp_priv,
						s_work);
	struct rvt_qp *qp = priv->owner;

	qib_do_send(qp);
}

/**
 * qib_do_send - perform a send on a QP
 * @qp: 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, two threads could send packets out of order.
 */
void qib_do_send(struct rvt_qp *qp)
{
	struct qib_qp_priv *priv = qp->priv;
	struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
	struct qib_pportdata *ppd = ppd_from_ibp(ibp);
	int (*make_req)(struct rvt_qp *qp, unsigned long *flags);
	unsigned long flags;

	if ((qp->ibqp.qp_type == IB_QPT_RC ||
	     qp->ibqp.qp_type == IB_QPT_UC) &&
	    (rdma_ah_get_dlid(&qp->remote_ah_attr) &
	     ~((1 << ppd->lmc) - 1)) == ppd->lid) {
		qib_ruc_loopback(qp);
		return;
	}

	if (qp->ibqp.qp_type == IB_QPT_RC)
		make_req = qib_make_rc_req;
	else if (qp->ibqp.qp_type == IB_QPT_UC)
		make_req = qib_make_uc_req;
	else
		make_req = qib_make_ud_req;

	spin_lock_irqsave(&qp->s_lock, flags);

	/* Return if we are already busy processing a work request. */
	if (!qib_send_ok(qp)) {
		spin_unlock_irqrestore(&qp->s_lock, flags);
		return;
	}

	qp->s_flags |= RVT_S_BUSY;

	do {
		/* Check for a constructed packet to be sent. */
		if (qp->s_hdrwords != 0) {
			spin_unlock_irqrestore(&qp->s_lock, flags);
			/*
			 * If the packet cannot be sent now, return and
			 * the send tasklet will be woken up later.
			 */
			if (qib_verbs_send(qp, priv->s_hdr, qp->s_hdrwords,
					   qp->s_cur_sge, qp->s_cur_size))
				return;
			/* Record that s_hdr is empty. */
			qp->s_hdrwords = 0;
			spin_lock_irqsave(&qp->s_lock, flags);
		}
	} while (make_req(qp, &flags));

	spin_unlock_irqrestore(&qp->s_lock, flags);
}

/*
 * This should be called with s_lock held.
 */
void qib_send_complete(struct rvt_qp *qp, struct rvt_swqe *wqe,
		       enum ib_wc_status status)
{
	u32 old_last, last;

	if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_OR_FLUSH_SEND))
		return;

	last = qp->s_last;
	old_last = last;
	if (++last >= qp->s_size)
		last = 0;
	qp->s_last = last;
	/* See post_send() */
	barrier();
	rvt_put_swqe(wqe);
	if (qp->ibqp.qp_type == IB_QPT_UD ||
	    qp->ibqp.qp_type == IB_QPT_SMI ||
	    qp->ibqp.qp_type == IB_QPT_GSI)
		atomic_dec(&ibah_to_rvtah(wqe->ud_wr.ah)->refcount);

	rvt_qp_swqe_complete(qp,
			     wqe,
			     ib_qib_wc_opcode[wqe->wr.opcode],
			     status);

	if (qp->s_acked == old_last)
		qp->s_acked = 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;
}