mthca_qp.c

linux内核源码

		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
			err = -EINVAL;
			*bad_wr = wr;
			goto out;
		}

		for (i = 0; i < wr->num_sge; ++i) {
			mthca_set_data_seg(wqe, wr->sg_list + i);
			wqe  += sizeof (struct mthca_data_seg);
			size += sizeof (struct mthca_data_seg) / 16;
		}

		qp->wrid[ind] = wr->wr_id;

		((struct mthca_next_seg *) prev_wqe)->nda_op =
			cpu_to_be32((ind << qp->rq.wqe_shift) | 1);
		wmb();
		((struct mthca_next_seg *) prev_wqe)->ee_nds =
			cpu_to_be32(MTHCA_NEXT_DBD | size);

		if (!nreq)
			size0 = size;

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

		++nreq;
		if (unlikely(nreq == MTHCA_TAVOR_MAX_WQES_PER_RECV_DB)) {
			nreq = 0;

			wmb();

			mthca_write64((qp->rq.next_ind << qp->rq.wqe_shift) | size0,
				      qp->qpn << 8, dev->kar + MTHCA_RECEIVE_DOORBELL,
				      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));

			qp->rq.next_ind = ind;
			qp->rq.head += MTHCA_TAVOR_MAX_WQES_PER_RECV_DB;
		}
	}

out:
	if (likely(nreq)) {
		wmb();

		mthca_write64((qp->rq.next_ind << qp->rq.wqe_shift) | size0,
			      qp->qpn << 8 | nreq, dev->kar + MTHCA_RECEIVE_DOORBELL,
			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
	}

	qp->rq.next_ind = ind;
	qp->rq.head    += nreq;

	/*
	 * Make sure doorbells don't leak out of RQ spinlock and reach
	 * the HCA out of order:
	 */
	mmiowb();

	spin_unlock_irqrestore(&qp->rq.lock, flags);
	return err;
}

int mthca_arbel_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);
	u32 dbhi;
	void *wqe;
	void *prev_wqe;
	unsigned long flags;
	int err = 0;
	int nreq;
	int i;
	int size;
	/*
	 * f0 and size0 are only used if nreq != 0, and they will
	 * always be initialized the first time through the main loop
	 * before nreq is incremented.  So nreq cannot become non-zero
	 * without initializing f0 and size0, and they are in fact
	 * never used uninitialized.
	 */
	int uninitialized_var(size0);
	u32 uninitialized_var(f0);
	int ind;
	u8 op0 = 0;

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

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

	ind = qp->sq.head & (qp->sq.max - 1);

	for (nreq = 0; wr; ++nreq, wr = wr->next) {
		if (unlikely(nreq == MTHCA_ARBEL_MAX_WQES_PER_SEND_DB)) {
			nreq = 0;

			dbhi = (MTHCA_ARBEL_MAX_WQES_PER_SEND_DB << 24) |
				((qp->sq.head & 0xffff) << 8) | f0 | op0;

			qp->sq.head += MTHCA_ARBEL_MAX_WQES_PER_SEND_DB;

			/*
			 * Make sure that descriptors are written before
			 * doorbell record.
			 */
			wmb();
			*qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);

			/*
			 * Make sure doorbell record is written before we
			 * write MMIO send doorbell.
			 */
			wmb();

			mthca_write64(dbhi, (qp->qpn << 8) | size0,
				      dev->kar + MTHCA_SEND_DOORBELL,
				      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
		}

		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)->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:
				set_raddr_seg(wqe, wr->wr.atomic.remote_addr,
					      wr->wr.atomic.rkey);
				wqe += sizeof (struct mthca_raddr_seg);

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

			case IB_WR_RDMA_READ:
			case IB_WR_RDMA_WRITE:
			case IB_WR_RDMA_WRITE_WITH_IMM:
				set_raddr_seg(wqe, wr->wr.rdma.remote_addr,
					      wr->wr.rdma.rkey);
				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:
				set_raddr_seg(wqe, wr->wr.rdma.remote_addr,
					      wr->wr.rdma.rkey);
				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:
			set_arbel_ud_seg(wqe, wr);
			wqe  += sizeof (struct mthca_arbel_ud_seg);
			size += sizeof (struct mthca_arbel_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) {
			mthca_set_data_seg(wqe, wr->sg_list + i);
			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(MTHCA_NEXT_DBD | size |
				    ((wr->send_flags & IB_SEND_FENCE) ?
				     MTHCA_NEXT_FENCE : 0));

		if (!nreq) {
			size0 = size;
			op0   = mthca_opcode[wr->opcode];
			f0    = wr->send_flags & IB_SEND_FENCE ?
				MTHCA_SEND_DOORBELL_FENCE : 0;
		}

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

out:
	if (likely(nreq)) {
		dbhi = (nreq << 24) | ((qp->sq.head & 0xffff) << 8) | f0 | op0;

		qp->sq.head += nreq;

		/*
		 * Make sure that descriptors are written before
		 * doorbell record.
		 */
		wmb();
		*qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);

		/*
		 * Make sure doorbell record is written before we
		 * write MMIO send doorbell.
		 */
		wmb();

		mthca_write64(dbhi, (qp->qpn << 8) | size0, dev->kar + MTHCA_SEND_DOORBELL,
			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
	}

	/*
	 * Make sure doorbells don't leak out of SQ spinlock and reach
	 * the HCA out of order:
	 */
	mmiowb();

	spin_unlock_irqrestore(&qp->sq.lock, flags);
	return err;
}

int mthca_arbel_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
			     struct ib_recv_wr **bad_wr)
{
	struct mthca_dev *dev = to_mdev(ibqp->device);
	struct mthca_qp *qp = to_mqp(ibqp);
	unsigned long flags;
	int err = 0;
	int nreq;
	int ind;
	int i;
	void *wqe;

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

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

	ind = qp->rq.head & (qp->rq.max - 1);

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

		wqe = get_recv_wqe(qp, ind);

		((struct mthca_next_seg *) wqe)->flags = 0;

		wqe += sizeof (struct mthca_next_seg);

		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
			err = -EINVAL;
			*bad_wr = wr;
			goto out;
		}

		for (i = 0; i < wr->num_sge; ++i) {
			mthca_set_data_seg(wqe, wr->sg_list + i);
			wqe += sizeof (struct mthca_data_seg);
		}

		if (i < qp->rq.max_gs)
			mthca_set_data_seg_inval(wqe);

		qp->wrid[ind] = wr->wr_id;

		++ind;
		if (unlikely(ind >= qp->rq.max))
			ind -= qp->rq.max;
	}
out:
	if (likely(nreq)) {
		qp->rq.head += nreq;

		/*
		 * Make sure that descriptors are written before
		 * doorbell record.
		 */
		wmb();
		*qp->rq.db = cpu_to_be32(qp->rq.head & 0xffff);
	}

	spin_unlock_irqrestore(&qp->rq.lock, flags);
	return err;
}

void mthca_free_err_wqe(struct mthca_dev *dev, struct mthca_qp *qp, int is_send,
			int index, int *dbd, __be32 *new_wqe)
{
	struct mthca_next_seg *next;

	/*
	 * For SRQs, all receive WQEs generate a CQE, so we're always
	 * at the end of the doorbell chain.
	 */
	if (qp->ibqp.srq && !is_send) {
		*new_wqe = 0;
		return;
	}

	if (is_send)
		next = get_send_wqe(qp, index);
	else
		next = get_recv_wqe(qp, index);

	*dbd = !!(next->ee_nds & cpu_to_be32(MTHCA_NEXT_DBD));
	if (next->ee_nds & cpu_to_be32(0x3f))
		*new_wqe = (next->nda_op & cpu_to_be32(~0x3f)) |
			(next->ee_nds & cpu_to_be32(0x3f));
	else
		*new_wqe = 0;
}

int mthca_init_qp_table(struct mthca_dev *dev)
{
	int err;
	u8 status;
	int i;

	spin_lock_init(&dev->qp_table.lock);

	/*
	 * We reserve 2 extra QPs per port for the special QPs.  The
	 * special QP for port 1 has to be even, so round up.
	 */
	dev->qp_table.sqp_start = (dev->limits.reserved_qps + 1) & ~1UL;
	err = mthca_alloc_init(&dev->qp_table.alloc,
			       dev->limits.num_qps,
			       (1 << 24) - 1,
			       dev->qp_table.sqp_start +
			       MTHCA_MAX_PORTS * 2);
	if (err)
		return err;

	err = mthca_array_init(&dev->qp_table.qp,
			       dev->limits.num_qps);
	if (err) {
		mthca_alloc_cleanup(&dev->qp_table.alloc);
		return err;
	}

	for (i = 0; i < 2; ++i) {
		err = mthca_CONF_SPECIAL_QP(dev, i ? IB_QPT_GSI : IB_QPT_SMI,
					    dev->qp_table.sqp_start + i * 2,
					    &status);
		if (err)
			goto err_out;
		if (status) {
			mthca_warn(dev, "CONF_SPECIAL_QP returned "
				   "status %02x, aborting.\n",
				   status);
			err = -EINVAL;
			goto err_out;
		}
	}
	return 0;

 err_out:
	for (i = 0; i < 2; ++i)
		mthca_CONF_SPECIAL_QP(dev, i, 0, &status);

	mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
	mthca_alloc_cleanup(&dev->qp_table.alloc);

	return err;
}

void mthca_cleanup_qp_table(struct mthca_dev *dev)
{
	int i;
	u8 status;

	for (i = 0; i < 2; ++i)
		mthca_CONF_SPECIAL_QP(dev, i, 0, &status);

	mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
	mthca_alloc_cleanup(&dev->qp_table.alloc);
}