mthca_qp.c

linux-2.6.15.6

	qp->atomic_rd_en = 0;
	qp->resp_depth   = 0;
	qp->sq_policy    = send_policy;
	mthca_wq_init(&qp->sq);
	mthca_wq_init(&qp->rq);

	ret = mthca_map_memfree(dev, qp);
	if (ret)
		return ret;

	ret = mthca_alloc_wqe_buf(dev, pd, qp);
	if (ret) {
		mthca_unmap_memfree(dev, qp);
		return ret;
	}

	mthca_adjust_qp_caps(dev, pd, qp);

	/*
	 * If this is a userspace QP, we're done now.  The doorbells
	 * will be allocated and buffers will be initialized in
	 * userspace.
	 */
	if (pd->ibpd.uobject)
		return 0;

	ret = mthca_alloc_memfree(dev, qp);
	if (ret) {
		mthca_free_wqe_buf(dev, qp);
		mthca_unmap_memfree(dev, qp);
		return ret;
	}

	if (mthca_is_memfree(dev)) {
		struct mthca_next_seg *next;
		struct mthca_data_seg *scatter;
		int size = (sizeof (struct mthca_next_seg) +
			    qp->rq.max_gs * sizeof (struct mthca_data_seg)) / 16;

		for (i = 0; i < qp->rq.max; ++i) {
			next = get_recv_wqe(qp, i);
			next->nda_op = cpu_to_be32(((i + 1) & (qp->rq.max - 1)) <<
						   qp->rq.wqe_shift);
			next->ee_nds = cpu_to_be32(size);

			for (scatter = (void *) (next + 1);
			     (void *) scatter < (void *) next + (1 << qp->rq.wqe_shift);
			     ++scatter)
				scatter->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
		}

		for (i = 0; i < qp->sq.max; ++i) {
			next = get_send_wqe(qp, i);
			next->nda_op = cpu_to_be32((((i + 1) & (qp->sq.max - 1)) <<
						    qp->sq.wqe_shift) +
						   qp->send_wqe_offset);
		}
	}

	qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
	qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);

	return 0;
}

static int mthca_set_qp_size(struct mthca_dev *dev, struct ib_qp_cap *cap,
			     struct mthca_pd *pd, struct mthca_qp *qp)
{
	int max_data_size = mthca_max_data_size(dev, qp, dev->limits.max_desc_sz);

	/* Sanity check QP size before proceeding */
	if (cap->max_send_wr  	 > dev->limits.max_wqes ||
	    cap->max_recv_wr  	 > dev->limits.max_wqes ||
	    cap->max_send_sge 	 > dev->limits.max_sg   ||
	    cap->max_recv_sge 	 > dev->limits.max_sg   ||
	    cap->max_inline_data > mthca_max_inline_data(pd, max_data_size))
		return -EINVAL;

	/*
	 * For MLX transport we need 2 extra S/G entries:
	 * one for the header and one for the checksum at the end
	 */
	if (qp->transport == MLX && cap->max_recv_sge + 2 > dev->limits.max_sg)
		return -EINVAL;

	if (mthca_is_memfree(dev)) {
		qp->rq.max = cap->max_recv_wr ?
			roundup_pow_of_two(cap->max_recv_wr) : 0;
		qp->sq.max = cap->max_send_wr ?
			roundup_pow_of_two(cap->max_send_wr) : 0;
	} else {
		qp->rq.max = cap->max_recv_wr;
		qp->sq.max = cap->max_send_wr;
	}

	qp->rq.max_gs = cap->max_recv_sge;
	qp->sq.max_gs = max_t(int, cap->max_send_sge,
			      ALIGN(cap->max_inline_data + MTHCA_INLINE_HEADER_SIZE,
				    MTHCA_INLINE_CHUNK_SIZE) /
			      sizeof (struct mthca_data_seg));

	return 0;
}

int mthca_alloc_qp(struct mthca_dev *dev,
		   struct mthca_pd *pd,
		   struct mthca_cq *send_cq,
		   struct mthca_cq *recv_cq,
		   enum ib_qp_type type,
		   enum ib_sig_type send_policy,
		   struct ib_qp_cap *cap,
		   struct mthca_qp *qp)
{
	int err;

	err = mthca_set_qp_size(dev, cap, pd, qp);
	if (err)
		return err;

	switch (type) {
	case IB_QPT_RC: qp->transport = RC; break;
	case IB_QPT_UC: qp->transport = UC; break;
	case IB_QPT_UD: qp->transport = UD; break;
	default: return -EINVAL;
	}

	qp->qpn = mthca_alloc(&dev->qp_table.alloc);
	if (qp->qpn == -1)
		return -ENOMEM;

	err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
				    send_policy, qp);
	if (err) {
		mthca_free(&dev->qp_table.alloc, qp->qpn);
		return err;
	}

	spin_lock_irq(&dev->qp_table.lock);
	mthca_array_set(&dev->qp_table.qp,
			qp->qpn & (dev->limits.num_qps - 1), qp);
	spin_unlock_irq(&dev->qp_table.lock);

	return 0;
}

int mthca_alloc_sqp(struct mthca_dev *dev,
		    struct mthca_pd *pd,
		    struct mthca_cq *send_cq,
		    struct mthca_cq *recv_cq,
		    enum ib_sig_type send_policy,
		    struct ib_qp_cap *cap,
		    int qpn,
		    int port,
		    struct mthca_sqp *sqp)
{
	u32 mqpn = qpn * 2 + dev->qp_table.sqp_start + port - 1;
	int err;

	err = mthca_set_qp_size(dev, cap, pd, &sqp->qp);
	if (err)
		return err;

	sqp->header_buf_size = sqp->qp.sq.max * MTHCA_UD_HEADER_SIZE;
	sqp->header_buf = dma_alloc_coherent(&dev->pdev->dev, sqp->header_buf_size,
					     &sqp->header_dma, GFP_KERNEL);
	if (!sqp->header_buf)
		return -ENOMEM;

	spin_lock_irq(&dev->qp_table.lock);
	if (mthca_array_get(&dev->qp_table.qp, mqpn))
		err = -EBUSY;
	else
		mthca_array_set(&dev->qp_table.qp, mqpn, sqp);
	spin_unlock_irq(&dev->qp_table.lock);

	if (err)
		goto err_out;

	sqp->port = port;
	sqp->qp.qpn       = mqpn;
	sqp->qp.transport = MLX;

	err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
				    send_policy, &sqp->qp);
	if (err)
		goto err_out_free;

	atomic_inc(&pd->sqp_count);

	return 0;

 err_out_free:
	/*
	 * Lock CQs here, so that CQ polling code can do QP lookup
	 * without taking a lock.
	 */
	spin_lock_irq(&send_cq->lock);
	if (send_cq != recv_cq)
		spin_lock(&recv_cq->lock);

	spin_lock(&dev->qp_table.lock);
	mthca_array_clear(&dev->qp_table.qp, mqpn);
	spin_unlock(&dev->qp_table.lock);

	if (send_cq != recv_cq)
		spin_unlock(&recv_cq->lock);
	spin_unlock_irq(&send_cq->lock);

 err_out:
	dma_free_coherent(&dev->pdev->dev, sqp->header_buf_size,
			  sqp->header_buf, sqp->header_dma);

	return err;
}

void mthca_free_qp(struct mthca_dev *dev,
		   struct mthca_qp *qp)
{
	u8 status;
	struct mthca_cq *send_cq;
	struct mthca_cq *recv_cq;

	send_cq = to_mcq(qp->ibqp.send_cq);
	recv_cq = to_mcq(qp->ibqp.recv_cq);

	/*
	 * Lock CQs here, so that CQ polling code can do QP lookup
	 * without taking a lock.
	 */
	spin_lock_irq(&send_cq->lock);
	if (send_cq != recv_cq)
		spin_lock(&recv_cq->lock);

	spin_lock(&dev->qp_table.lock);
	mthca_array_clear(&dev->qp_table.qp,
			  qp->qpn & (dev->limits.num_qps - 1));
	spin_unlock(&dev->qp_table.lock);

	if (send_cq != recv_cq)
		spin_unlock(&recv_cq->lock);
	spin_unlock_irq(&send_cq->lock);

	atomic_dec(&qp->refcount);
	wait_event(qp->wait, !atomic_read(&qp->refcount));

	if (qp->state != IB_QPS_RESET)
		mthca_MODIFY_QP(dev, MTHCA_TRANS_ANY2RST, qp->qpn, 0, NULL, 0, &status);

	/*
	 * If this is a userspace QP, the buffers, MR, CQs and so on
	 * will be cleaned up in userspace, so all we have to do is
	 * unref the mem-free tables and free the QPN in our table.
	 */
	if (!qp->ibqp.uobject) {
		mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq)->cqn, qp->qpn,
			       qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
		if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
			mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq)->cqn, qp->qpn,
				       qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);

		mthca_free_memfree(dev, qp);
		mthca_free_wqe_buf(dev, qp);
	}

	mthca_unmap_memfree(dev, qp);

	if (is_sqp(dev, qp)) {
		atomic_dec(&(to_mpd(qp->ibqp.pd)->sqp_count));
		dma_free_coherent(&dev->pdev->dev,
				  to_msqp(qp)->header_buf_size,
				  to_msqp(qp)->header_buf,
				  to_msqp(qp)->header_dma);
	} else
		mthca_free(&dev->qp_table.alloc, qp->qpn);
}

/* Create UD header for an MLX send and build a data segment for it */
static int build_mlx_header(struct mthca_dev *dev, struct mthca_sqp *sqp,
			    int ind, struct ib_send_wr *wr,
			    struct mthca_mlx_seg *mlx,
			    struct mthca_data_seg *data)
{
	int header_size;
	int err;
	u16 pkey;

	ib_ud_header_init(256, /* assume a MAD */
			  sqp->ud_header.grh_present,
			  &sqp->ud_header);

	err = mthca_read_ah(dev, to_mah(wr->wr.ud.ah), &sqp->ud_header);
	if (err)
		return err;
	mlx->flags &= ~cpu_to_be32(MTHCA_NEXT_SOLICIT | 1);
	mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MTHCA_MLX_VL15 : 0) |
				  (sqp->ud_header.lrh.destination_lid ==
				   IB_LID_PERMISSIVE ? MTHCA_MLX_SLR : 0) |
				  (sqp->ud_header.lrh.service_level << 8));
	mlx->rlid = sqp->ud_header.lrh.destination_lid;
	mlx->vcrc = 0;

	switch (wr->opcode) {
	case IB_WR_SEND:
		sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
		sqp->ud_header.immediate_present = 0;
		break;
	case IB_WR_SEND_WITH_IMM:
		sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
		sqp->ud_header.immediate_present = 1;
		sqp->ud_header.immediate_data = wr->imm_data;
		break;
	default:
		return -EINVAL;
	}

	sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 : 0;
	if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
		sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
	sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
	if (!sqp->qp.ibqp.qp_num)
		ib_get_cached_pkey(&dev->ib_dev, sqp->port,
				   sqp->pkey_index, &pkey);
	else
		ib_get_cached_pkey(&dev->ib_dev, sqp->port,
				   wr->wr.ud.pkey_index, &pkey);
	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
	sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
	sqp->ud_header.deth.qkey = cpu_to_be32(wr->wr.ud.remote_qkey & 0x80000000 ?
					       sqp->qkey : wr->wr.ud.remote_qkey);
	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);

	header_size = ib_ud_header_pack(&sqp->ud_header,
					sqp->header_buf +
					ind * MTHCA_UD_HEADER_SIZE);

	data->byte_count = cpu_to_be32(header_size);
	data->lkey       = cpu_to_be32(to_mpd(sqp->qp.ibqp.pd)->ntmr.ibmr.lkey);
	data->addr       = cpu_to_be64(sqp->header_dma +
				       ind * MTHCA_UD_HEADER_SIZE);

	return 0;
}

static inline int mthca_wq_overflow(struct mthca_wq *wq, int nreq,
				    struct ib_cq *ib_cq)
{
	unsigned cur;
	struct mthca_cq *cq;

	cur = wq->head - wq->tail;
	if (likely(cur + nreq < wq->max))
		return 0;

	cq = to_mcq(ib_cq);
	spin_lock(&cq->lock);
	cur = wq->head - wq->tail;
	spin_unlock(&cq->lock);

	return cur + nreq >= wq->max;
}

int mthca_tavor_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
			  struct ib_send_wr **bad_wr)
{
	struct mthca_dev *dev = to_mdev(ibqp->device);
	struct mthca_qp *qp = to_mqp(ibqp);
	void *wqe;
	void *prev_wqe;
	unsigned long flags;
	int err = 0;
	int nreq;
	int i;
	int size;
	int size0 = 0;
	u32 f0 = 0;
	int ind;
	u8 op0 = 0;

	spin_lock_irqsave(&qp->sq.lock, flags);

	/* XXX check that state is OK to post send */

	ind = qp->sq.next_ind;

	for (nreq = 0; wr; ++nreq, wr = wr->next) {
		if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
			mthca_err(dev, "SQ %06x full (%u head, %u tail,"
					" %d max, %d nreq)\n", qp->qpn,
					qp->sq.head, qp->sq.tail,
					qp->sq.max, nreq);
			err = -ENOMEM;
			*bad_wr = wr;
			goto out;
		}

		wqe = get_send_wqe(qp, ind);
		prev_wqe = qp->sq.last;
		qp->sq.last = wqe;

		((struct mthca_next_seg *) wqe)->nda_op = 0;
		((struct mthca_next_seg *) wqe)->ee_nds = 0;
		((struct mthca_next_seg *) wqe)->flags =
			((wr->send_flags & IB_SEND_SIGNALED) ?
			 cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
			((wr->send_flags & IB_SEND_SOLICITED) ?
			 cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0)   |
			cpu_to_be32(1);
		if (wr->opcode == IB_WR_SEND_WITH_IMM ||
		    wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
			((struct mthca_next_seg *) wqe)->imm = wr->imm_data;

		wqe += sizeof (struct mthca_next_seg);
		size = sizeof (struct mthca_next_seg) / 16;

		switch (qp->transport) {
		case RC:
			switch (wr->opcode) {
			case IB_WR_ATOMIC_CMP_AND_SWP:
			case IB_WR_ATOMIC_FETCH_AND_ADD:
				((struct mthca_raddr_seg *) wqe)->raddr =
					cpu_to_be64(wr->wr.atomic.remote_addr);
				((struct mthca_raddr_seg *) wqe)->rkey =
					cpu_to_be32(wr->wr.atomic.rkey);
				((struct mthca_raddr_seg *) wqe)->reserved = 0;

				wqe += sizeof (struct mthca_raddr_seg);

				if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
					((struct mthca_atomic_seg *) wqe)->swap_add =
						cpu_to_be64(wr->wr.atomic.swap);
					((struct mthca_atomic_seg *) wqe)->compare =
						cpu_to_be64(wr->wr.atomic.compare_add);
				} else {
					((struct mthca_atomic_seg *) wqe)->swap_add =
						cpu_to_be64(wr->wr.atomic.compare_add);
					((struct mthca_atomic_seg *) wqe)->compare = 0;
				}

				wqe += sizeof (struct mthca_atomic_seg);
				size += (sizeof (struct mthca_raddr_seg) +
					 sizeof (struct mthca_atomic_seg)) / 16;
				break;

			case IB_WR_RDMA_WRITE:
			case IB_WR_RDMA_WRITE_WITH_IMM:
			case IB_WR_RDMA_READ:
				((struct mthca_raddr_seg *) wqe)->raddr =
					cpu_to_be64(wr->wr.rdma.remote_addr);
				((struct mthca_raddr_seg *) wqe)->rkey =
					cpu_to_be32(wr->wr.rdma.rkey);
				((struct mthca_raddr_seg *) wqe)->reserved = 0;
				wqe += sizeof (struct mthca_raddr_seg);
				size += sizeof (struct mthca_raddr_seg) / 16;
				break;

			default:
				/* No extra segments required for sends */
				break;
			}

			break;

		case UC:
			switch (wr->opcode) {
			case IB_WR_RDMA_WRITE:
			case IB_WR_RDMA_WRITE_WITH_IMM:
				((struct mthca_raddr_seg *) wqe)->raddr =
					cpu_to_be64(wr->wr.rdma.remote_addr);
				((struct mthca_raddr_seg *) wqe)->rkey =
					cpu_to_be32(wr->wr.rdma.rkey);
				((struct mthca_raddr_seg *) wqe)->reserved = 0;
				wqe += sizeof (struct mthca_raddr_seg);
				size += sizeof (struct mthca_raddr_seg) / 16;
				break;

			default:
				/* No extra segments required for sends */
				break;
			}

			break;

		case UD:
			((struct mthca_tavor_ud_seg *) wqe)->lkey =
				cpu_to_be32(to_mah(wr->wr.ud.ah)->key);
			((struct mthca_tavor_ud_seg *) wqe)->av_addr =
				cpu_to_be64(to_mah(wr->wr.ud.ah)->avdma);
			((struct mthca_tavor_ud_seg *) wqe)->dqpn =
				cpu_to_be32(wr->wr.ud.remote_qpn);
			((struct mthca_tavor_ud_seg *) wqe)->qkey =
				cpu_to_be32(wr->wr.ud.remote_qkey);

			wqe += sizeof (struct mthca_tavor_ud_seg);
			size += sizeof (struct mthca_tavor_ud_seg) / 16;
			break;

		case MLX:
			err = build_mlx_header(dev, to_msqp(qp), ind, wr,
					       wqe - sizeof (struct mthca_next_seg),
					       wqe);
			if (err) {
				*bad_wr = wr;
				goto out;
			}
			wqe += sizeof (struct mthca_data_seg);
			size += sizeof (struct mthca_data_seg) / 16;
			break;
		}

		if (wr->num_sge > qp->sq.max_gs) {
			mthca_err(dev, "too many gathers\n");
			err = -EINVAL;
			*bad_wr = wr;
			goto out;
		}

		for (i = 0; i < wr->num_sge; ++i) {
			((struct mthca_data_seg *) wqe)->byte_count =
				cpu_to_be32(wr->sg_list[i].length);
			((struct mthca_data_seg *) wqe)->lkey =
				cpu_to_be32(wr->sg_list[i].lkey);
			((struct mthca_data_seg *) wqe)->addr =
				cpu_to_be64(wr->sg_list[i].addr);
			wqe += sizeof (struct mthca_data_seg);
			size += sizeof (struct mthca_data_seg) / 16;
		}

		/* Add one more inline data segment for ICRC */
		if (qp->transport == MLX) {
			((struct mthca_data_seg *) wqe)->byte_count =
				cpu_to_be32((1 << 31) | 4);
			((u32 *) wqe)[1] = 0;
			wqe += sizeof (struct mthca_data_seg);
			size += sizeof (struct mthca_data_seg) / 16;
		}

		qp->wrid[ind + qp->rq.max] = wr->wr_id;

		if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
			mthca_err(dev, "opcode invalid\n");
			err = -EINVAL;
			*bad_wr = wr;
			goto out;
		}

		((struct mthca_next_seg *) prev_wqe)->nda_op =
			cpu_to_be32(((ind << qp->sq.wqe_shift) +
				     qp->send_wqe_offset) |
				    mthca_opcode[wr->opcode]);
		wmb();
		((struct mthca_next_seg *) prev_wqe)->ee_nds =
			cpu_to_be32((size0 ? 0 : MTHCA_NEXT_DBD) | size);

		if (!size0) {
			size0 = size;
			op0   = mthca_opcode[wr->opcode];
		}

		++ind;
		if (unlikely(ind >= qp->sq.max))
			ind -= qp->sq.max;
	}

out: