[net-next-2.6.git] / drivers / infiniband / hw / ipath / ipath_ud.c

/*
 * 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
 * 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 <rdma/ib_smi.h>

#include "ipath_verbs.h"
#include "ipath_kernel.h"

/**
 * ipath_ud_loopback - handle send on loopback QPs
 * @sqp: the sending QP
 * @swqe: the send work request
 *
 * 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_swqe *swqe)
{
	struct ipath_ibdev *dev = to_idev(sqp->ibqp.device);
	struct ipath_qp *qp;
	struct ib_ah_attr *ah_attr;
	unsigned long flags;
	struct ipath_rq *rq;
	struct ipath_srq *srq;
	struct ipath_sge_state rsge;
	struct ipath_sge *sge;
	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, swqe->wr.wr.ud.remote_qpn);
	if (!qp || !(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_RECV_OK)) {
		dev->n_pkt_drops++;
		goto done;
	}

	rsge.sg_list = NULL;

	/*
	 * Check that the qkey matches (except for QP0, see 9.6.1.4.1).
	 * Qkeys with the high order bit set mean use the
	 * qkey from the QP context instead of the WR (see 10.2.5).
	 */
	if (unlikely(qp->ibqp.qp_num &&
		     ((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 drop;
	}

	/*
	 * A GRH is expected to preceed the data even if not
	 * 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.ex.imm_data = swqe->wr.ex.imm_data;
	}

	/*
	 * 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);
		handler = srq->ibsrq.event_handler;
		rq = &srq->rq;
	} else {
		srq = NULL;
		handler = NULL;
		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;
	/* 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) {
		spin_unlock_irqrestore(&rq->lock, flags);
		dev->n_pkt_drops++;
		goto drop;
	}
	if (++tail >= rq->size)
		tail = 0;
	wq->tail = tail;
	wc.wr_id = wqe->wr_id;
	if (handler) {
		u32 n;

		/*
		 * validate head pointer value and compute
		 * the number of remaining WQEs.
		 */
		n = wq->head;
		if (n >= rq->size)
			n = 0;
		if (n < tail)
			n += rq->size - tail;
		else
			n -= tail;
		if (n < srq->limit) {
			struct ib_event ev;

			srq->limit = 0;
			spin_unlock_irqrestore(&rq->lock, flags);
			ev.device = qp->ibqp.device;
			ev.element.srq = qp->ibqp.srq;
			ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
			handler(&ev, srq->ibsrq.srq_context);
		} else
			spin_unlock_irqrestore(&rq->lock, flags);
	} else
		spin_unlock_irqrestore(&rq->lock, flags);

	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;
	} else
		ipath_skip_sge(&rsge, sizeof(struct ib_grh));
	sge = swqe->sg_list;
	while (length) {
		u32 len = sge->length;

		if (len > length)
			len = length;
		if (len > sge->sge_length)
			len = sge->sge_length;
		BUG_ON(len == 0);
		ipath_copy_sge(&rsge, sge->vaddr, len);
		sge->vaddr += len;
		sge->length -= len;
		sge->sge_length -= len;
		if (sge->sge_length == 0) {
			if (--swqe->wr.num_sge)
				sge++;
		} else if (sge->length == 0 && sge->mr != NULL) {
			if (++sge->n >= IPATH_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;
		}
		length -= len;
	}
	wc.status = IB_WC_SUCCESS;
	wc.opcode = IB_WC_RECV;
	wc.qp = &qp->ibqp;
	wc.src_qp = sqp->ibqp.qp_num;
	/* XXX do we know which pkey matched? Only needed for GSI. */
	wc.pkey_index = 0;
	wc.slid = dev->dd->ipath_lid |
		(ah_attr->src_path_bits &
		 ((1 << dev->dd->ipath_lmc) - 1));
	wc.sl = ah_attr->sl;
	wc.dlid_path_bits =
		ah_attr->dlid & ((1 << dev->dd->ipath_lmc) - 1);
	wc.port_num = 1;
	/* Signal completion event if the solicited bit is set. */
	ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
		       swqe->wr.send_flags & IB_SEND_SOLICITED);
drop:
	kfree(rsge.sg_list);
	if (atomic_dec_and_test(&qp->refcount))
		wake_up(&qp->wait);
done:;
}

/**
 * ipath_make_ud_req - construct a UD request packet
 * @qp: the QP
 *
 * Return 1 if constructed; otherwise, return 0.
 */
int ipath_make_ud_req(struct ipath_qp *qp)
{
	struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
	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;
	u16 lrh0;
	u16 lid;
	int ret = 0;

	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;
	if (ah_attr->dlid >= IPATH_MULTICAST_LID_BASE) {
		if (ah_attr->dlid != IPATH_PERMISSIVE_LID)
			dev->n_multicast_xmit++;
		else
			dev->n_unicast_xmit++;
	} else {
		dev->n_unicast_xmit++;
		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;
		}
	}

	extra_bytes = -wqe->length & 3;
	nwords = (wqe->length + extra_bytes) >> 2;

	/* header size in 32-bit words LRH+BTH+DETH = (8+12+8)/4. */
	qp->s_hdrwords = 7;
	qp->s_cur_size = wqe->length;
	qp->s_cur_sge = &qp->s_sge;
	qp->s_dmult = ah_attr->static_rate;
	qp->s_wqe = wqe;
	qp->s_sge.sge = wqe->sg_list[0];
	qp->s_sge.sg_list = wqe->sg_list + 1;
	qp->s_sge.num_sge = wqe->wr.num_sge;

	if (ah_attr->ah_flags & IB_AH_GRH) {
		/* Header size in 32-bit words. */
		qp->s_hdrwords += ipath_make_grh(dev, &qp->s_hdr.u.l.grh,
						 &ah_attr->grh,
						 qp->s_hdrwords, nwords);
		lrh0 = IPATH_LRH_GRH;
		ohdr = &qp->s_hdr.u.l.oth;
		/*
		 * Don't worry about sending to locally attached multicast
		 * QPs.  It is unspecified by the spec. what happens.
		 */
	} else {
		/* Header size in 32-bit words. */
		lrh0 = IPATH_LRH_BTH;
		ohdr = &qp->s_hdr.u.oth;
	}
	if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
		qp->s_hdrwords++;
		ohdr->u.ud.imm_data = wqe->wr.ex.imm_data;
		bth0 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE << 24;
	} else
		bth0 = IB_OPCODE_UD_SEND_ONLY << 24;
	lrh0 |= ah_attr->sl << 4;
	if (qp->ibqp.qp_type == IB_QPT_SMI)
		lrh0 |= 0xF000;	/* Set VL (see ch. 13.5.3.1) */
	qp->s_hdr.lrh[0] = cpu_to_be16(lrh0);
	qp->s_hdr.lrh[1] = cpu_to_be16(ah_attr->dlid);	/* DEST LID */
	qp->s_hdr.lrh[2] = cpu_to_be16(qp->s_hdrwords + nwords +
					   SIZE_OF_CRC);
	lid = dev->dd->ipath_lid;
	if (lid) {
		lid |= ah_attr->src_path_bits &
			((1 << dev->dd->ipath_lmc) - 1);
		qp->s_hdr.lrh[3] = cpu_to_be16(lid);
	} else
		qp->s_hdr.lrh[3] = IB_LID_PERMISSIVE;
	if (wqe->wr.send_flags & IB_SEND_SOLICITED)
		bth0 |= 1 << 23;
	bth0 |= extra_bytes << 20;
	bth0 |= qp->ibqp.qp_type == IB_QPT_SMI ? IPATH_DEFAULT_P_KEY :
		ipath_get_pkey(dev->dd, qp->s_pkey_index);
	ohdr->bth[0] = cpu_to_be32(bth0);
	/*
	 * Use the multicast QP if the destination LID is a multicast LID.
	 */
	ohdr->bth[1] = ah_attr->dlid >= IPATH_MULTICAST_LID_BASE &&
		ah_attr->dlid != IPATH_PERMISSIVE_LID ?
		__constant_cpu_to_be32(IPATH_MULTICAST_QPN) :
		cpu_to_be32(wqe->wr.wr.ud.remote_qpn);
	ohdr->bth[2] = cpu_to_be32(qp->s_next_psn++ & IPATH_PSN_MASK);
	/*
	 * Qkeys with the high order bit set mean use the
	 * qkey from the QP context instead of the WR (see 10.2.5).
	 */
	ohdr->u.ud.deth[0] = cpu_to_be32((int)wqe->wr.wr.ud.remote_qkey < 0 ?
					 qp->qkey : wqe->wr.wr.ud.remote_qkey);
	ohdr->u.ud.deth[1] = cpu_to_be32(qp->ibqp.qp_num);

done:
	ret = 1;
	goto unlock;

bail:
	qp->s_flags &= ~IPATH_S_BUSY;
unlock:
	spin_unlock_irqrestore(&qp->s_lock, flags);
	return ret;
}

/**
 * ipath_ud_rcv - receive an incoming UD packet
 * @dev: the device the packet came in on
 * @hdr: the packet header
 * @has_grh: true if the packet has a GRH
 * @data: the packet data
 * @tlen: the packet length
 * @qp: the QP the packet came on
 *
 * This is called from ipath_qp_rcv() to process an incoming UD packet
 * for the given QP.
 * Called at interrupt level.
 */
void ipath_ud_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
		  int has_grh, void *data, u32 tlen, struct ipath_qp *qp)
{
	struct ipath_other_headers *ohdr;
	int opcode;
	u32 hdrsize;
	u32 pad;
	struct ib_wc wc;
	u32 qkey;
	u32 src_qp;
	u16 dlid;
	int header_in_data;

	/* Check for GRH */
	if (!has_grh) {
		ohdr = &hdr->u.oth;
		hdrsize = 8 + 12 + 8;	/* LRH + BTH + DETH */
		qkey = be32_to_cpu(ohdr->u.ud.deth[0]);
		src_qp = be32_to_cpu(ohdr->u.ud.deth[1]);
		header_in_data = 0;
	} else {
		ohdr = &hdr->u.l.oth;
		hdrsize = 8 + 40 + 12 + 8; /* LRH + GRH + BTH + DETH */
		/*
		 * The header with GRH is 68 bytes and the core driver sets
		 * the eager header buffer size to 56 bytes so the last 12
		 * bytes of the IB header is in the data buffer.
		 */
		header_in_data = dev->dd->ipath_rcvhdrentsize == 16;
		if (header_in_data) {
			qkey = be32_to_cpu(((__be32 *) data)[1]);
			src_qp = be32_to_cpu(((__be32 *) data)[2]);
			data += 12;
		} else {
			qkey = be32_to_cpu(ohdr->u.ud.deth[0]);
			src_qp = be32_to_cpu(ohdr->u.ud.deth[1]);
		}
	}
	src_qp &= IPATH_QPN_MASK;

	/*
	 * Check that the permissive LID is only used on QP0
	 * and the QKEY matches (see 9.6.1.4.1 and 9.6.1.5.1).
	 */
	if (qp->ibqp.qp_num) {
		if (unlikely(hdr->lrh[1] == IB_LID_PERMISSIVE ||
			     hdr->lrh[3] == IB_LID_PERMISSIVE)) {
			dev->n_pkt_drops++;
			goto bail;
		}
		if (unlikely(qkey != qp->qkey)) {
			/* XXX OK to lose a count once in a while. */
			dev->qkey_violations++;
			dev->n_pkt_drops++;
			goto bail;
		}
	} else if (hdr->lrh[1] == IB_LID_PERMISSIVE ||
		   hdr->lrh[3] == IB_LID_PERMISSIVE) {
		struct ib_smp *smp = (struct ib_smp *) data;

		if (smp->mgmt_class != IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
			dev->n_pkt_drops++;
			goto bail;
		}
	}

	/*
	 * The opcode is in the low byte when its in network order
	 * (top byte when in host order).
	 */
	opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
	if (qp->ibqp.qp_num > 1 &&
	    opcode == IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE) {
		if (header_in_data) {
			wc.ex.imm_data = *(__be32 *) data;
			data += sizeof(__be32);
		} else
			wc.ex.imm_data = ohdr->u.ud.imm_data;
		wc.wc_flags = IB_WC_WITH_IMM;
		hdrsize += sizeof(u32);
	} else if (opcode == IB_OPCODE_UD_SEND_ONLY) {
		wc.ex.imm_data = 0;
		wc.wc_flags = 0;
	} else {
		dev->n_pkt_drops++;
		goto bail;
	}

	/* Get the number of bytes the message was padded by. */
	pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
	if (unlikely(tlen < (hdrsize + pad + 4))) {
		/* Drop incomplete packets. */
		dev->n_pkt_drops++;
		goto bail;
	}
	tlen -= hdrsize + pad + 4;

	/* Drop invalid MAD packets (see 13.5.3.1). */
	if (unlikely((qp->ibqp.qp_num == 0 &&
		      (tlen != 256 ||
		       (be16_to_cpu(hdr->lrh[0]) >> 12) != 15)) ||
		     (qp->ibqp.qp_num == 1 &&
		      (tlen != 256 ||
		       (be16_to_cpu(hdr->lrh[0]) >> 12) == 15)))) {
		dev->n_pkt_drops++;
		goto bail;
	}

	/*
	 * A GRH is expected to preceed the data even if not
	 * present on the wire.
	 */
	wc.byte_len = tlen + sizeof(struct ib_grh);

	/*
	 * Get the next work request entry to find where to put the data.
	 */
	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.
		 * Otherwise, count them as buffer overruns since usually,
		 * the HW will be able to receive packets even if there are
		 * no QPs with posted receive buffers.
		 */
		if (qp->ibqp.qp_num == 0)
			dev->n_vl15_dropped++;
		else
			dev->rcv_errors++;
		goto bail;
	}
	/* Silently drop packets which are too big. */
	if (wc.byte_len > qp->r_len) {
		qp->r_flags |= IPATH_R_REUSE_SGE;
		dev->n_pkt_drops++;
		goto bail;
	}
	if (has_grh) {
		ipath_copy_sge(&qp->r_sge, &hdr->u.l.grh,
			       sizeof(struct ib_grh));
		wc.wc_flags |= IB_WC_GRH;
	} else
		ipath_skip_sge(&qp->r_sge, sizeof(struct ib_grh));
	ipath_copy_sge(&qp->r_sge, data,
		       wc.byte_len - sizeof(struct ib_grh));
	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;
	wc.vendor_err = 0;
	wc.qp = &qp->ibqp;
	wc.src_qp = src_qp;
	/* XXX do we know which pkey matched? Only needed for GSI. */
	wc.pkey_index = 0;
	wc.slid = be16_to_cpu(hdr->lrh[3]);
	wc.sl = (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF;
	dlid = be16_to_cpu(hdr->lrh[1]);
	/*
	 * Save the LMC lower bits if the destination LID is a unicast LID.
	 */
	wc.dlid_path_bits = dlid >= IPATH_MULTICAST_LID_BASE ? 0 :
		dlid & ((1 << dev->dd->ipath_lmc) - 1);
	wc.port_num = 1;
	/* Signal completion event if the solicited bit is set. */
	ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
		       (ohdr->bth[0] &
			__constant_cpu_to_be32(1 << 23)) != 0);

bail:;
}
Commit	Line	Data
74ed6b5e	1	/*
e7eacd36	2	* Copyright (c) 2006, 2007, 2008 QLogic Corporation. All rights reserved.
74ed6b5e BS	3	* Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
	4	*
	5	* This software is available to you under a choice of one of two
	6	* licenses. You may choose to be licensed under the terms of the GNU
	7	* General Public License (GPL) Version 2, available from the file
	8	* COPYING in the main directory of this source tree, or the
	9	* OpenIB.org BSD license below:
	10	*
	11	* Redistribution and use in source and binary forms, with or
	12	* without modification, are permitted provided that the following
	13	* conditions are met:
	14	*
	15	* - Redistributions of source code must retain the above
	16	* copyright notice, this list of conditions and the following
	17	* disclaimer.
	18	*
	19	* - Redistributions in binary form must reproduce the above
	20	* copyright notice, this list of conditions and the following
	21	* disclaimer in the documentation and/or other materials
	22	* provided with the distribution.
	23	*
	24	* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
	25	* EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
	26	* MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
	27	* NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
	28	* BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
	29	* ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
	30	* CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
	31	* SOFTWARE.
	32	*/
	33
	34	#include <rdma/ib_smi.h>
	35
	36	#include "ipath_verbs.h"
373d9915 RC	37	#include "ipath_kernel.h"
373d9915 RC	38
74ed6b5e BS	39	/**
74ed6b5e BS	40	* ipath_ud_loopback - handle send on loopback QPs
4ee97180 RC	41	* @sqp: the sending QP
4ee97180 RC	42	* @swqe: the send work request
74ed6b5e	43	*
4ee97180	44	* This is called from ipath_make_ud_req() to forward a WQE addressed
74ed6b5e	45	* to the same HCA.
373d9915 RC	46	* Note that the receive interrupt handler may be calling ipath_ud_rcv()
373d9915 RC	47	* while this is being called.
74ed6b5e	48	*/
4ee97180	49	static void ipath_ud_loopback(struct ipath_qp sqp, struct ipath_swqe swqe)
74ed6b5e BS	50	{
	51	struct ipath_ibdev *dev = to_idev(sqp->ibqp.device);
	52	struct ipath_qp *qp;
	53	struct ib_ah_attr *ah_attr;
	54	unsigned long flags;
	55	struct ipath_rq *rq;
	56	struct ipath_srq *srq;
	57	struct ipath_sge_state rsge;
	58	struct ipath_sge *sge;
373d9915	59	struct ipath_rwq *wq;
74ed6b5e	60	struct ipath_rwqe *wqe;
373d9915	61	void (handler)(struct ib_event , void *);
4ee97180	62	struct ib_wc wc;
373d9915 RC	63	u32 tail;
373d9915 RC	64	u32 rlen;
4ee97180	65	u32 length;
74ed6b5e	66
4ee97180	67	qp = ipath_lookup_qpn(&dev->qp_table, swqe->wr.wr.ud.remote_qpn);
e509be89	68	if (!qp \|\| !(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_RECV_OK)) {
4ee97180	69	dev->n_pkt_drops++;
e509be89	70	goto done;
4ee97180 RC	71	}
	72
	73	rsge.sg_list = NULL;
74ed6b5e BS	74
	75	/*
	76	* Check that the qkey matches (except for QP0, see 9.6.1.4.1).
	77	* Qkeys with the high order bit set mean use the
	78	* qkey from the QP context instead of the WR (see 10.2.5).
	79	*/
	80	if (unlikely(qp->ibqp.qp_num &&
4ee97180 RC	81	((int) swqe->wr.wr.ud.remote_qkey < 0 ?
4ee97180 RC	82	sqp->qkey : swqe->wr.wr.ud.remote_qkey) != qp->qkey)) {
74ed6b5e BS	83	/* XXX OK to lose a count once in a while. */
	84	dev->qkey_violations++;
	85	dev->n_pkt_drops++;
4ee97180	86	goto drop;
74ed6b5e BS	87	}
	88
	89	/*
	90	* A GRH is expected to preceed the data even if not
	91	* present on the wire.
	92	*/
4ee97180	93	length = swqe->length;
e509be89	94	memset(&wc, 0, sizeof wc);
4ee97180	95	wc.byte_len = length + sizeof(struct ib_grh);
74ed6b5e	96
4ee97180 RC	97	if (swqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
4ee97180 RC	98	wc.wc_flags = IB_WC_WITH_IMM;
00f7ec36	99	wc.ex.imm_data = swqe->wr.ex.imm_data;
74ed6b5e BS	100	}
	101
	102	/*
4ee97180 RC	103	* This would be a lot simpler if we could call ipath_get_rwqe()
	104	* but that uses state that the receive interrupt handler uses
	105	* so we would need to lock out receive interrupts while doing
	106	* local loopback.
74ed6b5e BS	107	*/
	108	if (qp->ibqp.srq) {
	109	srq = to_isrq(qp->ibqp.srq);
373d9915	110	handler = srq->ibsrq.event_handler;
74ed6b5e BS	111	rq = &srq->rq;
	112	} else {
	113	srq = NULL;
373d9915	114	handler = NULL;
74ed6b5e BS	115	rq = &qp->r_rq;
74ed6b5e BS	116	}
373d9915	117
4ee97180 RC	118	if (rq->max_sge > 1) {
	119	/*
	120	* XXX We could use GFP_KERNEL if ipath_do_send()
	121	* was always called from the tasklet instead of
	122	* from ipath_post_send().
	123	*/
	124	rsge.sg_list = kmalloc((rq->max_sge - 1) *
	125	sizeof(struct ipath_sge),
	126	GFP_ATOMIC);
	127	if (!rsge.sg_list) {
	128	dev->n_pkt_drops++;
	129	goto drop;
	130	}
	131	}
	132
	133	/*
	134	* Get the next work request entry to find where to put the data.
	135	* Note that it is safe to drop the lock after changing rq->tail
	136	* since ipath_post_receive() won't fill the empty slot.
	137	*/
74ed6b5e	138	spin_lock_irqsave(&rq->lock, flags);
373d9915 RC	139	wq = rq->wq;
373d9915 RC	140	tail = wq->tail;
4ee97180 RC	141	/* Validate tail before using it since it is user writable. */
	142	if (tail >= rq->size)
	143	tail = 0;
	144	if (unlikely(tail == wq->head)) {
	145	spin_unlock_irqrestore(&rq->lock, flags);
	146	dev->n_pkt_drops++;
	147	goto drop;
	148	}
	149	wqe = get_rwqe_ptr(rq, tail);
	150	if (!ipath_init_sge(qp, wqe, &rlen, &rsge)) {
	151	spin_unlock_irqrestore(&rq->lock, flags);
	152	dev->n_pkt_drops++;
	153	goto drop;
74ed6b5e BS	154	}
74ed6b5e BS	155	/* Silently drop packets which are too big. */
4ee97180	156	if (wc.byte_len > rlen) {
74ed6b5e BS	157	spin_unlock_irqrestore(&rq->lock, flags);
74ed6b5e BS	158	dev->n_pkt_drops++;
4ee97180	159	goto drop;
74ed6b5e	160	}
4ee97180 RC	161	if (++tail >= rq->size)
4ee97180 RC	162	tail = 0;
373d9915	163	wq->tail = tail;
4ee97180	164	wc.wr_id = wqe->wr_id;
373d9915	165	if (handler) {
74ed6b5e BS	166	u32 n;
74ed6b5e BS	167
373d9915 RC	168	/*
	169	* validate head pointer value and compute
	170	* the number of remaining WQEs.
	171	*/
	172	n = wq->head;
	173	if (n >= rq->size)
	174	n = 0;
	175	if (n < tail)
	176	n += rq->size - tail;
74ed6b5e	177	else
373d9915	178	n -= tail;
74ed6b5e BS	179	if (n < srq->limit) {
	180	struct ib_event ev;
	181
	182	srq->limit = 0;
	183	spin_unlock_irqrestore(&rq->lock, flags);
	184	ev.device = qp->ibqp.device;
	185	ev.element.srq = qp->ibqp.srq;
	186	ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
373d9915	187	handler(&ev, srq->ibsrq.srq_context);
74ed6b5e BS	188	} else
	189	spin_unlock_irqrestore(&rq->lock, flags);
	190	} else
	191	spin_unlock_irqrestore(&rq->lock, flags);
373d9915	192
4ee97180	193	ah_attr = &to_iah(swqe->wr.wr.ud.ah)->attr;
74ed6b5e BS	194	if (ah_attr->ah_flags & IB_AH_GRH) {
74ed6b5e BS	195	ipath_copy_sge(&rsge, &ah_attr->grh, sizeof(struct ib_grh));
4ee97180	196	wc.wc_flags \|= IB_WC_GRH;
74ed6b5e BS	197	} else
74ed6b5e BS	198	ipath_skip_sge(&rsge, sizeof(struct ib_grh));
4ee97180	199	sge = swqe->sg_list;
74ed6b5e BS	200	while (length) {
	201	u32 len = sge->length;
	202
	203	if (len > length)
	204	len = length;
30d149ab RC	205	if (len > sge->sge_length)
30d149ab RC	206	len = sge->sge_length;
74ed6b5e BS	207	BUG_ON(len == 0);
	208	ipath_copy_sge(&rsge, sge->vaddr, len);
	209	sge->vaddr += len;
	210	sge->length -= len;
	211	sge->sge_length -= len;
	212	if (sge->sge_length == 0) {
4ee97180 RC	213	if (--swqe->wr.num_sge)
4ee97180 RC	214	sge++;
74ed6b5e BS	215	} else if (sge->length == 0 && sge->mr != NULL) {
	216	if (++sge->n >= IPATH_SEGSZ) {
	217	if (++sge->m >= sge->mr->mapsz)
	218	break;
	219	sge->n = 0;
	220	}
	221	sge->vaddr =
	222	sge->mr->map[sge->m]->segs[sge->n].vaddr;
	223	sge->length =
	224	sge->mr->map[sge->m]->segs[sge->n].length;
	225	}
	226	length -= len;
	227	}
4ee97180 RC	228	wc.status = IB_WC_SUCCESS;
4ee97180 RC	229	wc.opcode = IB_WC_RECV;
4ee97180 RC	230	wc.qp = &qp->ibqp;
4ee97180 RC	231	wc.src_qp = sqp->ibqp.qp_num;
74ed6b5e	232	/* XXX do we know which pkey matched? Only needed for GSI. */
4ee97180 RC	233	wc.pkey_index = 0;
4ee97180 RC	234	wc.slid = dev->dd->ipath_lid \|
74ed6b5e	235	(ah_attr->src_path_bits &
542869a1	236	((1 << dev->dd->ipath_lmc) - 1));
4ee97180 RC	237	wc.sl = ah_attr->sl;
4ee97180 RC	238	wc.dlid_path_bits =
542869a1	239	ah_attr->dlid & ((1 << dev->dd->ipath_lmc) - 1);
4ee97180	240	wc.port_num = 1;
74ed6b5e	241	/* Signal completion event if the solicited bit is set. */
4ee97180 RC	242	ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
	243	swqe->wr.send_flags & IB_SEND_SOLICITED);
	244	drop:
373d9915	245	kfree(rsge.sg_list);
74ed6b5e BS	246	if (atomic_dec_and_test(&qp->refcount))
74ed6b5e BS	247	wake_up(&qp->wait);
e509be89	248	done:;
74ed6b5e BS	249	}
	250
	251	/**
4ee97180	252	* ipath_make_ud_req - construct a UD request packet
74ed6b5e	253	* @qp: the QP
74ed6b5e	254	*
4ee97180	255	* Return 1 if constructed; otherwise, return 0.
74ed6b5e	256	*/
4ee97180	257	int ipath_make_ud_req(struct ipath_qp *qp)
74ed6b5e BS	258	{
	259	struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
	260	struct ipath_other_headers *ohdr;
	261	struct ib_ah_attr *ah_attr;
4ee97180	262	struct ipath_swqe *wqe;
e509be89	263	unsigned long flags;
74ed6b5e	264	u32 nwords;
74ed6b5e BS	265	u32 extra_bytes;
	266	u32 bth0;
	267	u16 lrh0;
	268	u16 lid;
4ee97180	269	int ret = 0;
74ed6b5e	270
e509be89 RC	271	spin_lock_irqsave(&qp->s_lock, flags);
	272
	273	if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_NEXT_SEND_OK)) {
	274	if (!(ib_ipath_state_ops[qp->state] & IPATH_FLUSH_SEND))
	275	goto bail;
	276	/* We are in the error state, flush the work request. */
	277	if (qp->s_last == qp->s_head)
	278	goto bail;
	279	/* If DMAs are in progress, we can't flush immediately. */
	280	if (atomic_read(&qp->s_dma_busy)) {
	281	qp->s_flags \|= IPATH_S_WAIT_DMA;
	282	goto bail;
	283	}
	284	wqe = get_swqe_ptr(qp, qp->s_last);
	285	ipath_send_complete(qp, wqe, IB_WC_WR_FLUSH_ERR);
	286	goto done;
	287	}
74ed6b5e	288
4ee97180	289	if (qp->s_cur == qp->s_head)
6ce73b07	290	goto bail;
6ce73b07	291
4ee97180	292	wqe = get_swqe_ptr(qp, qp->s_cur);
e509be89 RC	293	if (++qp->s_cur >= qp->s_size)
e509be89 RC	294	qp->s_cur = 0;
74ed6b5e BS	295
74ed6b5e BS	296	/* Construct the header. */
4ee97180	297	ah_attr = &to_iah(wqe->wr.wr.ud.ah)->attr;
27b678dd BS	298	if (ah_attr->dlid >= IPATH_MULTICAST_LID_BASE) {
27b678dd BS	299	if (ah_attr->dlid != IPATH_PERMISSIVE_LID)
74ed6b5e BS	300	dev->n_multicast_xmit++;
	301	else
	302	dev->n_unicast_xmit++;
	303	} else {
	304	dev->n_unicast_xmit++;
e509be89	305	lid = ah_attr->dlid & ~((1 << dev->dd->ipath_lmc) - 1);
34b2aafe	306	if (unlikely(lid == dev->dd->ipath_lid)) {
e509be89 RC	307	/*
	308	* If DMAs are in progress, we can't generate
	309	* a completion for the loopback packet since
	310	* it would be out of order.
	311	* XXX Instead of waiting, we could queue a
	312	* zero length descriptor so we get a callback.
	313	*/
	314	if (atomic_read(&qp->s_dma_busy)) {
	315	qp->s_flags \|= IPATH_S_WAIT_DMA;
	316	goto bail;
	317	}
	318	spin_unlock_irqrestore(&qp->s_lock, flags);
4ee97180	319	ipath_ud_loopback(qp, wqe);
e509be89 RC	320	spin_lock_irqsave(&qp->s_lock, flags);
e509be89 RC	321	ipath_send_complete(qp, wqe, IB_WC_SUCCESS);
74ed6b5e BS	322	goto done;
	323	}
	324	}
4ee97180 RC	325
	326	extra_bytes = -wqe->length & 3;
	327	nwords = (wqe->length + extra_bytes) >> 2;
	328
	329	/* header size in 32-bit words LRH+BTH+DETH = (8+12+8)/4. */
	330	qp->s_hdrwords = 7;
4ee97180 RC	331	qp->s_cur_size = wqe->length;
4ee97180 RC	332	qp->s_cur_sge = &qp->s_sge;
124b4dcb	333	qp->s_dmult = ah_attr->static_rate;
4ee97180 RC	334	qp->s_wqe = wqe;
	335	qp->s_sge.sge = wqe->sg_list[0];
	336	qp->s_sge.sg_list = wqe->sg_list + 1;
	337	qp->s_sge.num_sge = wqe->wr.num_sge;
	338
74ed6b5e BS	339	if (ah_attr->ah_flags & IB_AH_GRH) {
74ed6b5e BS	340	/* Header size in 32-bit words. */
4ee97180 RC	341	qp->s_hdrwords += ipath_make_grh(dev, &qp->s_hdr.u.l.grh,
	342	&ah_attr->grh,
	343	qp->s_hdrwords, nwords);
27b678dd	344	lrh0 = IPATH_LRH_GRH;
74ed6b5e	345	ohdr = &qp->s_hdr.u.l.oth;
74ed6b5e BS	346	/*
	347	* Don't worry about sending to locally attached multicast
	348	* QPs. It is unspecified by the spec. what happens.
	349	*/
	350	} else {
	351	/* Header size in 32-bit words. */
27b678dd	352	lrh0 = IPATH_LRH_BTH;
74ed6b5e BS	353	ohdr = &qp->s_hdr.u.oth;
74ed6b5e BS	354	}
4ee97180	355	if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
4e1e93a4	356	qp->s_hdrwords++;
0f39cf3d	357	ohdr->u.ud.imm_data = wqe->wr.ex.imm_data;
74ed6b5e	358	bth0 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE << 24;
4ee97180	359	} else
74ed6b5e	360	bth0 = IB_OPCODE_UD_SEND_ONLY << 24;
74ed6b5e BS	361	lrh0 \|= ah_attr->sl << 4;
	362	if (qp->ibqp.qp_type == IB_QPT_SMI)
	363	lrh0 \|= 0xF000; /* Set VL (see ch. 13.5.3.1) */
	364	qp->s_hdr.lrh[0] = cpu_to_be16(lrh0);
	365	qp->s_hdr.lrh[1] = cpu_to_be16(ah_attr->dlid); /* DEST LID */
4ee97180 RC	366	qp->s_hdr.lrh[2] = cpu_to_be16(qp->s_hdrwords + nwords +
4ee97180 RC	367	SIZE_OF_CRC);
34b2aafe	368	lid = dev->dd->ipath_lid;
74ed6b5e BS	369	if (lid) {
74ed6b5e BS	370	lid \|= ah_attr->src_path_bits &
542869a1	371	((1 << dev->dd->ipath_lmc) - 1);
74ed6b5e BS	372	qp->s_hdr.lrh[3] = cpu_to_be16(lid);
	373	} else
	374	qp->s_hdr.lrh[3] = IB_LID_PERMISSIVE;
4ee97180	375	if (wqe->wr.send_flags & IB_SEND_SOLICITED)
74ed6b5e BS	376	bth0 \|= 1 << 23;
74ed6b5e BS	377	bth0 \|= extra_bytes << 20;
27b678dd	378	bth0 \|= qp->ibqp.qp_type == IB_QPT_SMI ? IPATH_DEFAULT_P_KEY :
34b2aafe	379	ipath_get_pkey(dev->dd, qp->s_pkey_index);
74ed6b5e BS	380	ohdr->bth[0] = cpu_to_be32(bth0);
	381	/*
	382	* Use the multicast QP if the destination LID is a multicast LID.
	383	*/
27b678dd BS	384	ohdr->bth[1] = ah_attr->dlid >= IPATH_MULTICAST_LID_BASE &&
	385	ah_attr->dlid != IPATH_PERMISSIVE_LID ?
	386	__constant_cpu_to_be32(IPATH_MULTICAST_QPN) :
4ee97180	387	cpu_to_be32(wqe->wr.wr.ud.remote_qpn);
27b678dd	388	ohdr->bth[2] = cpu_to_be32(qp->s_next_psn++ & IPATH_PSN_MASK);
74ed6b5e BS	389	/*
	390	* Qkeys with the high order bit set mean use the
	391	* qkey from the QP context instead of the WR (see 10.2.5).
	392	*/
4ee97180 RC	393	ohdr->u.ud.deth[0] = cpu_to_be32((int)wqe->wr.wr.ud.remote_qkey < 0 ?
4ee97180 RC	394	qp->qkey : wqe->wr.wr.ud.remote_qkey);
74ed6b5e	395	ohdr->u.ud.deth[1] = cpu_to_be32(qp->ibqp.qp_num);
74ed6b5e BS	396
74ed6b5e BS	397	done:
4ee97180	398	ret = 1;
e509be89	399	goto unlock;
74ed6b5e BS	400
74ed6b5e BS	401	bail:
e509be89 RC	402	qp->s_flags &= ~IPATH_S_BUSY;
	403	unlock:
	404	spin_unlock_irqrestore(&qp->s_lock, flags);
74ed6b5e BS	405	return ret;
	406	}
	407
	408	/**
	409	* ipath_ud_rcv - receive an incoming UD packet
	410	* @dev: the device the packet came in on
	411	* @hdr: the packet header
	412	* @has_grh: true if the packet has a GRH
	413	* @data: the packet data
	414	* @tlen: the packet length
	415	* @qp: the QP the packet came on
	416	*
	417	* This is called from ipath_qp_rcv() to process an incoming UD packet
	418	* for the given QP.
	419	* Called at interrupt level.
	420	*/
	421	void ipath_ud_rcv(struct ipath_ibdev dev, struct ipath_ib_header hdr,
	422	int has_grh, void data, u32 tlen, struct ipath_qp qp)
	423	{
	424	struct ipath_other_headers *ohdr;
	425	int opcode;
	426	u32 hdrsize;
	427	u32 pad;
74ed6b5e BS	428	struct ib_wc wc;
	429	u32 qkey;
	430	u32 src_qp;
74ed6b5e BS	431	u16 dlid;
	432	int header_in_data;
	433
	434	/* Check for GRH */
	435	if (!has_grh) {
	436	ohdr = &hdr->u.oth;
	437	hdrsize = 8 + 12 + 8; /* LRH + BTH + DETH */
	438	qkey = be32_to_cpu(ohdr->u.ud.deth[0]);
	439	src_qp = be32_to_cpu(ohdr->u.ud.deth[1]);
	440	header_in_data = 0;
	441	} else {
	442	ohdr = &hdr->u.l.oth;
	443	hdrsize = 8 + 40 + 12 + 8; /* LRH + GRH + BTH + DETH */
	444	/*
	445	* The header with GRH is 68 bytes and the core driver sets
	446	* the eager header buffer size to 56 bytes so the last 12
	447	* bytes of the IB header is in the data buffer.
	448	*/
34b2aafe	449	header_in_data = dev->dd->ipath_rcvhdrentsize == 16;
74ed6b5e BS	450	if (header_in_data) {
	451	qkey = be32_to_cpu(((__be32 *) data)[1]);
	452	src_qp = be32_to_cpu(((__be32 *) data)[2]);
	453	data += 12;
	454	} else {
	455	qkey = be32_to_cpu(ohdr->u.ud.deth[0]);
	456	src_qp = be32_to_cpu(ohdr->u.ud.deth[1]);
	457	}
	458	}
27b678dd	459	src_qp &= IPATH_QPN_MASK;
74ed6b5e BS	460
	461	/*
	462	* Check that the permissive LID is only used on QP0
	463	* and the QKEY matches (see 9.6.1.4.1 and 9.6.1.5.1).
	464	*/
	465	if (qp->ibqp.qp_num) {
	466	if (unlikely(hdr->lrh[1] == IB_LID_PERMISSIVE \|\|
	467	hdr->lrh[3] == IB_LID_PERMISSIVE)) {
	468	dev->n_pkt_drops++;
	469	goto bail;
	470	}
	471	if (unlikely(qkey != qp->qkey)) {
	472	/* XXX OK to lose a count once in a while. */
	473	dev->qkey_violations++;
	474	dev->n_pkt_drops++;
	475	goto bail;
	476	}
	477	} else if (hdr->lrh[1] == IB_LID_PERMISSIVE \|\|
	478	hdr->lrh[3] == IB_LID_PERMISSIVE) {
	479	struct ib_smp smp = (struct ib_smp ) data;
	480
	481	if (smp->mgmt_class != IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
	482	dev->n_pkt_drops++;
	483	goto bail;
	484	}
	485	}
	486
0a69631b RC	487	/*
	488	* The opcode is in the low byte when its in network order
	489	* (top byte when in host order).
	490	*/
	491	opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
	492	if (qp->ibqp.qp_num > 1 &&
	493	opcode == IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE) {
	494	if (header_in_data) {
00f7ec36	495	wc.ex.imm_data = (__be32 ) data;
0a69631b RC	496	data += sizeof(__be32);
0a69631b RC	497	} else
00f7ec36	498	wc.ex.imm_data = ohdr->u.ud.imm_data;
0a69631b RC	499	wc.wc_flags = IB_WC_WITH_IMM;
	500	hdrsize += sizeof(u32);
	501	} else if (opcode == IB_OPCODE_UD_SEND_ONLY) {
00f7ec36	502	wc.ex.imm_data = 0;
0a69631b RC	503	wc.wc_flags = 0;
	504	} else {
	505	dev->n_pkt_drops++;
	506	goto bail;
	507	}
	508
74ed6b5e BS	509	/* Get the number of bytes the message was padded by. */
	510	pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
	511	if (unlikely(tlen < (hdrsize + pad + 4))) {
	512	/* Drop incomplete packets. */
	513	dev->n_pkt_drops++;
	514	goto bail;
	515	}
	516	tlen -= hdrsize + pad + 4;
	517
	518	/* Drop invalid MAD packets (see 13.5.3.1). */
	519	if (unlikely((qp->ibqp.qp_num == 0 &&
	520	(tlen != 256 \|\|
	521	(be16_to_cpu(hdr->lrh[0]) >> 12) != 15)) \|\|
	522	(qp->ibqp.qp_num == 1 &&
	523	(tlen != 256 \|\|
	524	(be16_to_cpu(hdr->lrh[0]) >> 12) == 15)))) {
	525	dev->n_pkt_drops++;
	526	goto bail;
	527	}
	528
	529	/*
	530	* A GRH is expected to preceed the data even if not
	531	* present on the wire.
	532	*/
	533	wc.byte_len = tlen + sizeof(struct ib_grh);
	534
74ed6b5e BS	535	/*
74ed6b5e BS	536	* Get the next work request entry to find where to put the data.
74ed6b5e	537	*/
e509be89 RC	538	if (qp->r_flags & IPATH_R_REUSE_SGE)
e509be89 RC	539	qp->r_flags &= ~IPATH_R_REUSE_SGE;
373d9915	540	else if (!ipath_get_rwqe(qp, 0)) {
fba75200 BS	541	/*
	542	* Count VL15 packets dropped due to no receive buffer.
	543	* Otherwise, count them as buffer overruns since usually,
	544	* the HW will be able to receive packets even if there are
	545	* no QPs with posted receive buffers.
	546	*/
	547	if (qp->ibqp.qp_num == 0)
	548	dev->n_vl15_dropped++;
	549	else
	550	dev->rcv_errors++;
74ed6b5e BS	551	goto bail;
	552	}
	553	/* Silently drop packets which are too big. */
373d9915	554	if (wc.byte_len > qp->r_len) {
e509be89	555	qp->r_flags \|= IPATH_R_REUSE_SGE;
74ed6b5e BS	556	dev->n_pkt_drops++;
	557	goto bail;
	558	}
74ed6b5e BS	559	if (has_grh) {
	560	ipath_copy_sge(&qp->r_sge, &hdr->u.l.grh,
	561	sizeof(struct ib_grh));
	562	wc.wc_flags \|= IB_WC_GRH;
	563	} else
	564	ipath_skip_sge(&qp->r_sge, sizeof(struct ib_grh));
	565	ipath_copy_sge(&qp->r_sge, data,
	566	wc.byte_len - sizeof(struct ib_grh));
e509be89 RC	567	if (!test_and_clear_bit(IPATH_R_WRID_VALID, &qp->r_aflags))
e509be89 RC	568	goto bail;
373d9915	569	wc.wr_id = qp->r_wr_id;
74ed6b5e BS	570	wc.status = IB_WC_SUCCESS;
	571	wc.opcode = IB_WC_RECV;
	572	wc.vendor_err = 0;
062dbb69	573	wc.qp = &qp->ibqp;
74ed6b5e BS	574	wc.src_qp = src_qp;
	575	/* XXX do we know which pkey matched? Only needed for GSI. */
	576	wc.pkey_index = 0;
	577	wc.slid = be16_to_cpu(hdr->lrh[3]);
	578	wc.sl = (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF;
	579	dlid = be16_to_cpu(hdr->lrh[1]);
	580	/*
	581	* Save the LMC lower bits if the destination LID is a unicast LID.
	582	*/
27b678dd	583	wc.dlid_path_bits = dlid >= IPATH_MULTICAST_LID_BASE ? 0 :
542869a1	584	dlid & ((1 << dev->dd->ipath_lmc) - 1);
4ee97180	585	wc.port_num = 1;
74ed6b5e BS	586	/* Signal completion event if the solicited bit is set. */
	587	ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
	588	(ohdr->bth[0] &
	589	__constant_cpu_to_be32(1 << 23)) != 0);
	590
	591	bail:;
	592	}