IB/mlx4: Initialize SR-IOV IB support for slaves in master context

Allocate SR-IOV paravirtualization resources and MAD demuxing contexts
on the master.

This has two parts.  The first part is to initialize the structures to
contain the contexts.  This is done at master startup time in
mlx4_ib_init_sriov().

The second part is to actually create the tunneling resources required
on the master to support a slave.  This is performed the master
detects that a slave has started up (MLX4_DEV_EVENT_SLAVE_INIT event
generated when a slave initializes its comm channel).

For the master, there is no such startup event, so it creates its own
tunneling resources when it starts up.  In addition, the master also
creates the real special QPs.  The ib_core layer on the master causes
creation of proxy special QPs, since the master is also
paravirtualized at the ib_core layer.

Signed-off-by: Jack Morgenstein <jackm@dev.mellanox.co.il>
Signed-off-by: Roland Dreier <roland@purestorage.com>

6 files changed, 798 insertions, 8 deletions

diff --git a/drivers/infiniband/hw/mlx4/mad.c b/drivers/infiniband/hw/mlx4/mad.c
index 9c2ae7efd00f..e98849338a94 100644
--- a/drivers/infiniband/hw/mlx4/mad.c
+++ b/drivers/infiniband/hw/mlx4/mad.c
@@ -44,6 +44,35 @@ enum {
 	MLX4_IB_VENDOR_CLASS2 = 0xa
 };
 
+#define MLX4_TUN_SEND_WRID_SHIFT 34
+#define MLX4_TUN_QPN_SHIFT 32
+#define MLX4_TUN_WRID_RECV (((u64) 1) << MLX4_TUN_SEND_WRID_SHIFT)
+#define MLX4_TUN_SET_WRID_QPN(a) (((u64) ((a) & 0x3)) << MLX4_TUN_QPN_SHIFT)
+
+#define MLX4_TUN_IS_RECV(a)  (((a) >>  MLX4_TUN_SEND_WRID_SHIFT) & 0x1)
+#define MLX4_TUN_WRID_QPN(a) (((a) >> MLX4_TUN_QPN_SHIFT) & 0x3)
+
+struct mlx4_mad_rcv_buf {
+	struct ib_grh grh;
+	u8 payload[256];
+} __packed;
+
+struct mlx4_mad_snd_buf {
+	u8 payload[256];
+} __packed;
+
+struct mlx4_tunnel_mad {
+	struct ib_grh grh;
+	struct mlx4_ib_tunnel_header hdr;
+	struct ib_mad mad;
+} __packed;
+
+struct mlx4_rcv_tunnel_mad {
+	struct mlx4_rcv_tunnel_hdr hdr;
+	struct ib_grh grh;
+	struct ib_mad mad;
+} __packed;
+
 int mlx4_MAD_IFC(struct mlx4_ib_dev *dev, int ignore_mkey, int ignore_bkey,
 		 int port, struct ib_wc *in_wc, struct ib_grh *in_grh,
 		 void *in_mad, void *response_mad)
@@ -516,3 +545,658 @@ void mlx4_ib_dispatch_event(struct mlx4_ib_dev *dev, u8 port_num,
 
 	ib_dispatch_event(&event);
 }
+
+static void mlx4_ib_tunnel_comp_handler(struct ib_cq *cq, void *arg)
+{
+	unsigned long flags;
+	struct mlx4_ib_demux_pv_ctx *ctx = cq->cq_context;
+	struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
+	spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
+	if (!dev->sriov.is_going_down && ctx->state == DEMUX_PV_STATE_ACTIVE)
+		queue_work(ctx->wq, &ctx->work);
+	spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
+}
+
+static int mlx4_ib_post_pv_qp_buf(struct mlx4_ib_demux_pv_ctx *ctx,
+				  struct mlx4_ib_demux_pv_qp *tun_qp,
+				  int index)
+{
+	struct ib_sge sg_list;
+	struct ib_recv_wr recv_wr, *bad_recv_wr;
+	int size;
+
+	size = (tun_qp->qp->qp_type == IB_QPT_UD) ?
+		sizeof (struct mlx4_tunnel_mad) : sizeof (struct mlx4_mad_rcv_buf);
+
+	sg_list.addr = tun_qp->ring[index].map;
+	sg_list.length = size;
+	sg_list.lkey = ctx->mr->lkey;
+
+	recv_wr.next = NULL;
+	recv_wr.sg_list = &sg_list;
+	recv_wr.num_sge = 1;
+	recv_wr.wr_id = (u64) index | MLX4_TUN_WRID_RECV |
+		MLX4_TUN_SET_WRID_QPN(tun_qp->proxy_qpt);
+	ib_dma_sync_single_for_device(ctx->ib_dev, tun_qp->ring[index].map,
+				      size, DMA_FROM_DEVICE);
+	return ib_post_recv(tun_qp->qp, &recv_wr, &bad_recv_wr);
+}
+
+static int mlx4_ib_alloc_pv_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
+				 enum ib_qp_type qp_type, int is_tun)
+{
+	int i;
+	struct mlx4_ib_demux_pv_qp *tun_qp;
+	int rx_buf_size, tx_buf_size;
+
+	if (qp_type > IB_QPT_GSI)
+		return -EINVAL;
+
+	tun_qp = &ctx->qp[qp_type];
+
+	tun_qp->ring = kzalloc(sizeof (struct mlx4_ib_buf) * MLX4_NUM_TUNNEL_BUFS,
+			       GFP_KERNEL);
+	if (!tun_qp->ring)
+		return -ENOMEM;
+
+	tun_qp->tx_ring = kcalloc(MLX4_NUM_TUNNEL_BUFS,
+				  sizeof (struct mlx4_ib_tun_tx_buf),
+				  GFP_KERNEL);
+	if (!tun_qp->tx_ring) {
+		kfree(tun_qp->ring);
+		tun_qp->ring = NULL;
+		return -ENOMEM;
+	}
+
+	if (is_tun) {
+		rx_buf_size = sizeof (struct mlx4_tunnel_mad);
+		tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
+	} else {
+		rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
+		tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
+	}
+
+	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
+		tun_qp->ring[i].addr = kmalloc(rx_buf_size, GFP_KERNEL);
+		if (!tun_qp->ring[i].addr)
+			goto err;
+		tun_qp->ring[i].map = ib_dma_map_single(ctx->ib_dev,
+							tun_qp->ring[i].addr,
+							rx_buf_size,
+							DMA_FROM_DEVICE);
+	}
+
+	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
+		tun_qp->tx_ring[i].buf.addr =
+			kmalloc(tx_buf_size, GFP_KERNEL);
+		if (!tun_qp->tx_ring[i].buf.addr)
+			goto tx_err;
+		tun_qp->tx_ring[i].buf.map =
+			ib_dma_map_single(ctx->ib_dev,
+					  tun_qp->tx_ring[i].buf.addr,
+					  tx_buf_size,
+					  DMA_TO_DEVICE);
+		tun_qp->tx_ring[i].ah = NULL;
+	}
+	spin_lock_init(&tun_qp->tx_lock);
+	tun_qp->tx_ix_head = 0;
+	tun_qp->tx_ix_tail = 0;
+	tun_qp->proxy_qpt = qp_type;
+
+	return 0;
+
+tx_err:
+	while (i > 0) {
+		--i;
+		ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
+				    tx_buf_size, DMA_TO_DEVICE);
+		kfree(tun_qp->tx_ring[i].buf.addr);
+	}
+	kfree(tun_qp->tx_ring);
+	tun_qp->tx_ring = NULL;
+	i = MLX4_NUM_TUNNEL_BUFS;
+err:
+	while (i > 0) {
+		--i;
+		ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
+				    rx_buf_size, DMA_FROM_DEVICE);
+		kfree(tun_qp->ring[i].addr);
+	}
+	kfree(tun_qp->ring);
+	tun_qp->ring = NULL;
+	return -ENOMEM;
+}
+
+static void mlx4_ib_free_pv_qp_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
+				     enum ib_qp_type qp_type, int is_tun)
+{
+	int i;
+	struct mlx4_ib_demux_pv_qp *tun_qp;
+	int rx_buf_size, tx_buf_size;
+
+	if (qp_type > IB_QPT_GSI)
+		return;
+
+	tun_qp = &ctx->qp[qp_type];
+	if (is_tun) {
+		rx_buf_size = sizeof (struct mlx4_tunnel_mad);
+		tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
+	} else {
+		rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
+		tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
+	}
+
+
+	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
+		ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
+				    rx_buf_size, DMA_FROM_DEVICE);
+		kfree(tun_qp->ring[i].addr);
+	}
+
+	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
+		ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
+				    tx_buf_size, DMA_TO_DEVICE);
+		kfree(tun_qp->tx_ring[i].buf.addr);
+		if (tun_qp->tx_ring[i].ah)
+			ib_destroy_ah(tun_qp->tx_ring[i].ah);
+	}
+	kfree(tun_qp->tx_ring);
+	kfree(tun_qp->ring);
+}
+
+static void mlx4_ib_tunnel_comp_worker(struct work_struct *work)
+{
+	/* dummy until next patch in series */
+}
+
+static void pv_qp_event_handler(struct ib_event *event, void *qp_context)
+{
+	struct mlx4_ib_demux_pv_ctx *sqp = qp_context;
+
+	/* It's worse than that! He's dead, Jim! */
+	pr_err("Fatal error (%d) on a MAD QP on port %d\n",
+	       event->event, sqp->port);
+}
+
+static int create_pv_sqp(struct mlx4_ib_demux_pv_ctx *ctx,
+			    enum ib_qp_type qp_type, int create_tun)
+{
+	int i, ret;
+	struct mlx4_ib_demux_pv_qp *tun_qp;
+	struct mlx4_ib_qp_tunnel_init_attr qp_init_attr;
+	struct ib_qp_attr attr;
+	int qp_attr_mask_INIT;
+
+	if (qp_type > IB_QPT_GSI)
+		return -EINVAL;
+
+	tun_qp = &ctx->qp[qp_type];
+
+	memset(&qp_init_attr, 0, sizeof qp_init_attr);
+	qp_init_attr.init_attr.send_cq = ctx->cq;
+	qp_init_attr.init_attr.recv_cq = ctx->cq;
+	qp_init_attr.init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
+	qp_init_attr.init_attr.cap.max_send_wr = MLX4_NUM_TUNNEL_BUFS;
+	qp_init_attr.init_attr.cap.max_recv_wr = MLX4_NUM_TUNNEL_BUFS;
+	qp_init_attr.init_attr.cap.max_send_sge = 1;
+	qp_init_attr.init_attr.cap.max_recv_sge = 1;
+	if (create_tun) {
+		qp_init_attr.init_attr.qp_type = IB_QPT_UD;
+		qp_init_attr.init_attr.create_flags = MLX4_IB_SRIOV_TUNNEL_QP;
+		qp_init_attr.port = ctx->port;
+		qp_init_attr.slave = ctx->slave;
+		qp_init_attr.proxy_qp_type = qp_type;
+		qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX |
+			   IB_QP_QKEY | IB_QP_PORT;
+	} else {
+		qp_init_attr.init_attr.qp_type = qp_type;
+		qp_init_attr.init_attr.create_flags = MLX4_IB_SRIOV_SQP;
+		qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_QKEY;
+	}
+	qp_init_attr.init_attr.port_num = ctx->port;
+	qp_init_attr.init_attr.qp_context = ctx;
+	qp_init_attr.init_attr.event_handler = pv_qp_event_handler;
+	tun_qp->qp = ib_create_qp(ctx->pd, &qp_init_attr.init_attr);
+	if (IS_ERR(tun_qp->qp)) {
+		ret = PTR_ERR(tun_qp->qp);
+		tun_qp->qp = NULL;
+		pr_err("Couldn't create %s QP (%d)\n",
+		       create_tun ? "tunnel" : "special", ret);
+		return ret;
+	}
+
+	memset(&attr, 0, sizeof attr);
+	attr.qp_state = IB_QPS_INIT;
+	attr.pkey_index =
+		to_mdev(ctx->ib_dev)->pkeys.virt2phys_pkey[ctx->slave][ctx->port - 1][0];
+	attr.qkey = IB_QP1_QKEY;
+	attr.port_num = ctx->port;
+	ret = ib_modify_qp(tun_qp->qp, &attr, qp_attr_mask_INIT);
+	if (ret) {
+		pr_err("Couldn't change %s qp state to INIT (%d)\n",
+		       create_tun ? "tunnel" : "special", ret);
+		goto err_qp;
+	}
+	attr.qp_state = IB_QPS_RTR;
+	ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE);
+	if (ret) {
+		pr_err("Couldn't change %s qp state to RTR (%d)\n",
+		       create_tun ? "tunnel" : "special", ret);
+		goto err_qp;
+	}
+	attr.qp_state = IB_QPS_RTS;
+	attr.sq_psn = 0;
+	ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE | IB_QP_SQ_PSN);
+	if (ret) {
+		pr_err("Couldn't change %s qp state to RTS (%d)\n",
+		       create_tun ? "tunnel" : "special", ret);
+		goto err_qp;
+	}
+
+	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
+		ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp, i);
+		if (ret) {
+			pr_err(" mlx4_ib_post_pv_buf error"
+			       " (err = %d, i = %d)\n", ret, i);
+			goto err_qp;
+		}
+	}
+	return 0;
+
+err_qp:
+	ib_destroy_qp(tun_qp->qp);
+	tun_qp->qp = NULL;
+	return ret;
+}
+
+/*
+ * IB MAD completion callback for real SQPs
+ */
+static void mlx4_ib_sqp_comp_worker(struct work_struct *work)
+{
+	/* dummy until next patch in series */
+}
+
+static int alloc_pv_object(struct mlx4_ib_dev *dev, int slave, int port,
+			       struct mlx4_ib_demux_pv_ctx **ret_ctx)
+{
+	struct mlx4_ib_demux_pv_ctx *ctx;
+
+	*ret_ctx = NULL;
+	ctx = kzalloc(sizeof (struct mlx4_ib_demux_pv_ctx), GFP_KERNEL);
+	if (!ctx) {
+		pr_err("failed allocating pv resource context "
+		       "for port %d, slave %d\n", port, slave);
+		return -ENOMEM;
+	}
+
+	ctx->ib_dev = &dev->ib_dev;
+	ctx->port = port;
+	ctx->slave = slave;
+	*ret_ctx = ctx;
+	return 0;
+}
+
+static void free_pv_object(struct mlx4_ib_dev *dev, int slave, int port)
+{
+	if (dev->sriov.demux[port - 1].tun[slave]) {
+		kfree(dev->sriov.demux[port - 1].tun[slave]);
+		dev->sriov.demux[port - 1].tun[slave] = NULL;
+	}
+}
+
+static int create_pv_resources(struct ib_device *ibdev, int slave, int port,
+			       int create_tun, struct mlx4_ib_demux_pv_ctx *ctx)
+{
+	int ret, cq_size;
+
+	ctx->state = DEMUX_PV_STATE_STARTING;
+	/* have QP0 only on port owner, and only if link layer is IB */
+	if (ctx->slave == mlx4_master_func_num(to_mdev(ctx->ib_dev)->dev) &&
+	    rdma_port_get_link_layer(ibdev, ctx->port) == IB_LINK_LAYER_INFINIBAND)
+		ctx->has_smi = 1;
+
+	if (ctx->has_smi) {
+		ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_SMI, create_tun);
+		if (ret) {
+			pr_err("Failed allocating qp0 tunnel bufs (%d)\n", ret);
+			goto err_out;
+		}
+	}
+
+	ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_GSI, create_tun);
+	if (ret) {
+		pr_err("Failed allocating qp1 tunnel bufs (%d)\n", ret);
+		goto err_out_qp0;
+	}
+
+	cq_size = 2 * MLX4_NUM_TUNNEL_BUFS;
+	if (ctx->has_smi)
+		cq_size *= 2;
+
+	ctx->cq = ib_create_cq(ctx->ib_dev, mlx4_ib_tunnel_comp_handler,
+			       NULL, ctx, cq_size, 0);
+	if (IS_ERR(ctx->cq)) {
+		ret = PTR_ERR(ctx->cq);
+		pr_err("Couldn't create tunnel CQ (%d)\n", ret);
+		goto err_buf;
+	}
+
+	ctx->pd = ib_alloc_pd(ctx->ib_dev);
+	if (IS_ERR(ctx->pd)) {
+		ret = PTR_ERR(ctx->pd);
+		pr_err("Couldn't create tunnel PD (%d)\n", ret);
+		goto err_cq;
+	}
+
+	ctx->mr = ib_get_dma_mr(ctx->pd, IB_ACCESS_LOCAL_WRITE);
+	if (IS_ERR(ctx->mr)) {
+		ret = PTR_ERR(ctx->mr);
+		pr_err("Couldn't get tunnel DMA MR (%d)\n", ret);
+		goto err_pd;
+	}
+
+	if (ctx->has_smi) {
+		ret = create_pv_sqp(ctx, IB_QPT_SMI, create_tun);
+		if (ret) {
+			pr_err("Couldn't create %s QP0 (%d)\n",
+			       create_tun ? "tunnel for" : "",  ret);
+			goto err_mr;
+		}
+	}
+
+	ret = create_pv_sqp(ctx, IB_QPT_GSI, create_tun);
+	if (ret) {
+		pr_err("Couldn't create %s QP1 (%d)\n",
+		       create_tun ? "tunnel for" : "",  ret);
+		goto err_qp0;
+	}
+
+	if (create_tun)
+		INIT_WORK(&ctx->work, mlx4_ib_tunnel_comp_worker);
+	else
+		INIT_WORK(&ctx->work, mlx4_ib_sqp_comp_worker);
+
+	ctx->wq = to_mdev(ibdev)->sriov.demux[port - 1].wq;
+
+	ret = ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
+	if (ret) {
+		pr_err("Couldn't arm tunnel cq (%d)\n", ret);
+		goto err_wq;
+	}
+	ctx->state = DEMUX_PV_STATE_ACTIVE;
+	return 0;
+
+err_wq:
+	ctx->wq = NULL;
+	ib_destroy_qp(ctx->qp[1].qp);
+	ctx->qp[1].qp = NULL;
+
+
+err_qp0:
+	if (ctx->has_smi)
+		ib_destroy_qp(ctx->qp[0].qp);
+	ctx->qp[0].qp = NULL;
+
+err_mr:
+	ib_dereg_mr(ctx->mr);
+	ctx->mr = NULL;
+
+err_pd:
+	ib_dealloc_pd(ctx->pd);
+	ctx->pd = NULL;
+
+err_cq:
+	ib_destroy_cq(ctx->cq);
+	ctx->cq = NULL;
+
+err_buf:
+	mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, create_tun);
+
+err_out_qp0:
+	if (ctx->has_smi)
+		mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, create_tun);
+err_out:
+	ctx->state = DEMUX_PV_STATE_DOWN;
+	return ret;
+}
+
+static void destroy_pv_resources(struct mlx4_ib_dev *dev, int slave, int port,
+				 struct mlx4_ib_demux_pv_ctx *ctx, int flush)
+{
+	if (!ctx)
+		return;
+	if (ctx->state > DEMUX_PV_STATE_DOWN) {
+		ctx->state = DEMUX_PV_STATE_DOWNING;
+		if (flush)
+			flush_workqueue(ctx->wq);
+		if (ctx->has_smi) {
+			ib_destroy_qp(ctx->qp[0].qp);
+			ctx->qp[0].qp = NULL;
+			mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, 1);
+		}
+		ib_destroy_qp(ctx->qp[1].qp);
+		ctx->qp[1].qp = NULL;
+		mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, 1);
+		ib_dereg_mr(ctx->mr);
+		ctx->mr = NULL;
+		ib_dealloc_pd(ctx->pd);
+		ctx->pd = NULL;
+		ib_destroy_cq(ctx->cq);
+		ctx->cq = NULL;
+		ctx->state = DEMUX_PV_STATE_DOWN;
+	}
+}
+
+static int mlx4_ib_tunnels_update(struct mlx4_ib_dev *dev, int slave,
+				  int port, int do_init)
+{
+	int ret = 0;
+
+	if (!do_init) {
+		/* for master, destroy real sqp resources */
+		if (slave == mlx4_master_func_num(dev->dev))
+			destroy_pv_resources(dev, slave, port,
+					     dev->sriov.sqps[port - 1], 1);
+		/* destroy the tunnel qp resources */
+		destroy_pv_resources(dev, slave, port,
+				     dev->sriov.demux[port - 1].tun[slave], 1);
+		return 0;
+	}
+
+	/* create the tunnel qp resources */
+	ret = create_pv_resources(&dev->ib_dev, slave, port, 1,
+				  dev->sriov.demux[port - 1].tun[slave]);
+
+	/* for master, create the real sqp resources */
+	if (!ret && slave == mlx4_master_func_num(dev->dev))
+		ret = create_pv_resources(&dev->ib_dev, slave, port, 0,
+					  dev->sriov.sqps[port - 1]);
+	return ret;
+}
+
+void mlx4_ib_tunnels_update_work(struct work_struct *work)
+{
+	struct mlx4_ib_demux_work *dmxw;
+
+	dmxw = container_of(work, struct mlx4_ib_demux_work, work);
+	mlx4_ib_tunnels_update(dmxw->dev, dmxw->slave, (int) dmxw->port,
+			       dmxw->do_init);
+	kfree(dmxw);
+	return;
+}
+
+static int mlx4_ib_alloc_demux_ctx(struct mlx4_ib_dev *dev,
+				       struct mlx4_ib_demux_ctx *ctx,
+				       int port)
+{
+	char name[12];
+	int ret = 0;
+	int i;
+
+	ctx->tun = kcalloc(dev->dev->caps.sqp_demux,
+			   sizeof (struct mlx4_ib_demux_pv_ctx *), GFP_KERNEL);
+	if (!ctx->tun)
+		return -ENOMEM;
+
+	ctx->dev = dev;
+	ctx->port = port;
+	ctx->ib_dev = &dev->ib_dev;
+
+	for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
+		ret = alloc_pv_object(dev, i, port, &ctx->tun[i]);
+		if (ret) {
+			ret = -ENOMEM;
+			goto err_wq;
+		}
+	}
+
+	snprintf(name, sizeof name, "mlx4_ibt%d", port);
+	ctx->wq = create_singlethread_workqueue(name);
+	if (!ctx->wq) {
+		pr_err("Failed to create tunnelling WQ for port %d\n", port);
+		ret = -ENOMEM;
+		goto err_wq;
+	}
+
+	snprintf(name, sizeof name, "mlx4_ibud%d", port);
+	ctx->ud_wq = create_singlethread_workqueue(name);
+	if (!ctx->ud_wq) {
+		pr_err("Failed to create up/down WQ for port %d\n", port);
+		ret = -ENOMEM;
+		goto err_udwq;
+	}
+
+	return 0;
+
+err_udwq:
+	destroy_workqueue(ctx->wq);
+	ctx->wq = NULL;
+
+err_wq:
+	for (i = 0; i < dev->dev->caps.sqp_demux; i++)
+		free_pv_object(dev, i, port);
+	kfree(ctx->tun);
+	ctx->tun = NULL;
+	return ret;
+}
+
+static void mlx4_ib_free_sqp_ctx(struct mlx4_ib_demux_pv_ctx *sqp_ctx)
+{
+	if (sqp_ctx->state > DEMUX_PV_STATE_DOWN) {
+		sqp_ctx->state = DEMUX_PV_STATE_DOWNING;
+		flush_workqueue(sqp_ctx->wq);
+		if (sqp_ctx->has_smi) {
+			ib_destroy_qp(sqp_ctx->qp[0].qp);
+			sqp_ctx->qp[0].qp = NULL;
+			mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_SMI, 0);
+		}
+		ib_destroy_qp(sqp_ctx->qp[1].qp);
+		sqp_ctx->qp[1].qp = NULL;
+		mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_GSI, 0);
+		ib_dereg_mr(sqp_ctx->mr);
+		sqp_ctx->mr = NULL;
+		ib_dealloc_pd(sqp_ctx->pd);
+		sqp_ctx->pd = NULL;
+		ib_destroy_cq(sqp_ctx->cq);
+		sqp_ctx->cq = NULL;
+		sqp_ctx->state = DEMUX_PV_STATE_DOWN;
+	}
+}
+
+static void mlx4_ib_free_demux_ctx(struct mlx4_ib_demux_ctx *ctx)
+{
+	int i;
+	if (ctx) {
+		struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
+		for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
+			if (!ctx->tun[i])
+				continue;
+			if (ctx->tun[i]->state > DEMUX_PV_STATE_DOWN)
+				ctx->tun[i]->state = DEMUX_PV_STATE_DOWNING;
+		}
+		flush_workqueue(ctx->wq);
+		for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
+			destroy_pv_resources(dev, i, ctx->port, ctx->tun[i], 0);
+			free_pv_object(dev, i, ctx->port);
+		}
+		kfree(ctx->tun);
+		destroy_workqueue(ctx->ud_wq);
+		destroy_workqueue(ctx->wq);
+	}
+}
+
+static void mlx4_ib_master_tunnels(struct mlx4_ib_dev *dev, int do_init)
+{
+	int i;
+
+	if (!mlx4_is_master(dev->dev))
+		return;
+	/* initialize or tear down tunnel QPs for the master */
+	for (i = 0; i < dev->dev->caps.num_ports; i++)
+		mlx4_ib_tunnels_update(dev, mlx4_master_func_num(dev->dev), i + 1, do_init);
+	return;
+}
+
+int mlx4_ib_init_sriov(struct mlx4_ib_dev *dev)
+{
+	int i = 0;
+	int err;
+
+	if (!mlx4_is_mfunc(dev->dev))
+		return 0;
+
+	dev->sriov.is_going_down = 0;
+	spin_lock_init(&dev->sriov.going_down_lock);
+
+	mlx4_ib_warn(&dev->ib_dev, "multi-function enabled\n");
+
+	if (mlx4_is_slave(dev->dev)) {
+		mlx4_ib_warn(&dev->ib_dev, "operating in qp1 tunnel mode\n");
+		return 0;
+	}
+
+	mlx4_ib_warn(&dev->ib_dev, "initializing demux service for %d qp1 clients\n",
+		     dev->dev->caps.sqp_demux);
+	for (i = 0; i < dev->num_ports; i++) {
+		err = alloc_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1,
+				      &dev->sriov.sqps[i]);
+		if (err)
+			goto demux_err;
+		err = mlx4_ib_alloc_demux_ctx(dev, &dev->sriov.demux[i], i + 1);
+		if (err)
+			goto demux_err;
+	}
+	mlx4_ib_master_tunnels(dev, 1);
+	return 0;
+
+demux_err:
+	while (i > 0) {
+		free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
+		mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
+		--i;
+	}
+
+	return err;
+}
+
+void mlx4_ib_close_sriov(struct mlx4_ib_dev *dev)
+{
+	int i;
+	unsigned long flags;
+
+	if (!mlx4_is_mfunc(dev->dev))
+		return;
+
+	spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
+	dev->sriov.is_going_down = 1;
+	spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
+	if (mlx4_is_master(dev->dev))
+		for (i = 0; i < dev->num_ports; i++) {
+			flush_workqueue(dev->sriov.demux[i].ud_wq);
+			mlx4_ib_free_sqp_ctx(dev->sriov.sqps[i]);
+			kfree(dev->sriov.sqps[i]);
+			dev->sriov.sqps[i] = NULL;
+			mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
+		}
+}
diff --git a/drivers/infiniband/hw/mlx4/main.c b/drivers/infiniband/hw/mlx4/main.c
index cc05579ebce7..3f7f77f93a1c 100644
--- a/drivers/infiniband/hw/mlx4/main.c
+++ b/drivers/infiniband/hw/mlx4/main.c
@@ -1357,11 +1357,14 @@ static void *mlx4_ib_add(struct mlx4_dev *dev)
 	if (mlx4_ib_mad_init(ibdev))
 		goto err_reg;
 
+	if (mlx4_ib_init_sriov(ibdev))
+		goto err_mad;
+
 	if (dev->caps.flags & MLX4_DEV_CAP_FLAG_IBOE && !iboe->nb.notifier_call) {
 		iboe->nb.notifier_call = mlx4_ib_netdev_event;
 		err = register_netdevice_notifier(&iboe->nb);
 		if (err)
-			goto err_reg;
+			goto err_sriov;
 	}
 
 	for (j = 0; j < ARRAY_SIZE(mlx4_class_attributes); ++j) {
@@ -1379,6 +1382,12 @@ err_notif:
 		pr_warn("failure unregistering notifier\n");
 	flush_workqueue(wq);
 
+err_sriov:
+	mlx4_ib_close_sriov(ibdev);
+
+err_mad:
+	mlx4_ib_mad_cleanup(ibdev);
+
 err_reg:
 	ib_unregister_device(&ibdev->ib_dev);
 
@@ -1407,6 +1416,7 @@ static void mlx4_ib_remove(struct mlx4_dev *dev, void *ibdev_ptr)
 	struct mlx4_ib_dev *ibdev = ibdev_ptr;
 	int p;
 
+	mlx4_ib_close_sriov(ibdev);
 	mlx4_ib_mad_cleanup(ibdev);
 	ib_unregister_device(&ibdev->ib_dev);
 	if (ibdev->iboe.nb.notifier_call) {
@@ -1428,6 +1438,51 @@ static void mlx4_ib_remove(struct mlx4_dev *dev, void *ibdev_ptr)
 	ib_dealloc_device(&ibdev->ib_dev);
 }
 
+static void do_slave_init(struct mlx4_ib_dev *ibdev, int slave, int do_init)
+{
+	struct mlx4_ib_demux_work **dm = NULL;
+	struct mlx4_dev *dev = ibdev->dev;
+	int i;
+	unsigned long flags;
+
+	if (!mlx4_is_master(dev))
+		return;
+
+	dm = kcalloc(dev->caps.num_ports, sizeof *dm, GFP_ATOMIC);
+	if (!dm) {
+		pr_err("failed to allocate memory for tunneling qp update\n");
+		goto out;
+	}
+
+	for (i = 0; i < dev->caps.num_ports; i++) {
+		dm[i] = kmalloc(sizeof (struct mlx4_ib_demux_work), GFP_ATOMIC);
+		if (!dm[i]) {
+			pr_err("failed to allocate memory for tunneling qp update work struct\n");
+			for (i = 0; i < dev->caps.num_ports; i++) {
+				if (dm[i])
+					kfree(dm[i]);
+			}
+			goto out;
+		}
+	}
+	/* initialize or tear down tunnel QPs for the slave */
+	for (i = 0; i < dev->caps.num_ports; i++) {
+		INIT_WORK(&dm[i]->work, mlx4_ib_tunnels_update_work);
+		dm[i]->port = i + 1;
+		dm[i]->slave = slave;
+		dm[i]->do_init = do_init;
+		dm[i]->dev = ibdev;
+		spin_lock_irqsave(&ibdev->sriov.going_down_lock, flags);
+		if (!ibdev->sriov.is_going_down)
+			queue_work(ibdev->sriov.demux[i].ud_wq, &dm[i]->work);
+		spin_unlock_irqrestore(&ibdev->sriov.going_down_lock, flags);
+	}
+out:
+	if (dm)
+		kfree(dm);
+	return;
+}
+
 static void mlx4_ib_event(struct mlx4_dev *dev, void *ibdev_ptr,
 			  enum mlx4_dev_event event, unsigned long param)
 {
@@ -1435,22 +1490,23 @@ static void mlx4_ib_event(struct mlx4_dev *dev, void *ibdev_ptr,
 	struct mlx4_ib_dev *ibdev = to_mdev((struct ib_device *) ibdev_ptr);
 	struct mlx4_eqe *eqe = NULL;
 	struct ib_event_work *ew;
-	int port = 0;
+	int p = 0;
 
 	if (event == MLX4_DEV_EVENT_PORT_MGMT_CHANGE)
 		eqe = (struct mlx4_eqe *)param;
 	else
-		port = (u8)param;
-
-	if (port > ibdev->num_ports)
-		return;
+		p = (int) param;
 
 	switch (event) {
 	case MLX4_DEV_EVENT_PORT_UP:
+		if (p > ibdev->num_ports)
+			return;
 		ibev.event = IB_EVENT_PORT_ACTIVE;
 		break;
 
 	case MLX4_DEV_EVENT_PORT_DOWN:
+		if (p > ibdev->num_ports)
+			return;
 		ibev.event = IB_EVENT_PORT_ERR;
 		break;
 
@@ -1472,12 +1528,22 @@ static void mlx4_ib_event(struct mlx4_dev *dev, void *ibdev_ptr,
 		handle_port_mgmt_change_event(&ew->work);
 		return;
 
+	case MLX4_DEV_EVENT_SLAVE_INIT:
+		/* here, p is the slave id */
+		do_slave_init(ibdev, p, 1);
+		return;
+
+	case MLX4_DEV_EVENT_SLAVE_SHUTDOWN:
+		/* here, p is the slave id */
+		do_slave_init(ibdev, p, 0);
+		return;
+
 	default:
 		return;
 	}
 
 	ibev.device	      = ibdev_ptr;
-	ibev.element.port_num = port;
+	ibev.element.port_num = (u8) p;
 
 	ib_dispatch_event(&ibev);
 }
diff --git a/drivers/infiniband/hw/mlx4/mlx4_ib.h b/drivers/infiniband/hw/mlx4/mlx4_ib.h
index 1248d576b031..137941d79870 100644
--- a/drivers/infiniband/hw/mlx4/mlx4_ib.h
+++ b/drivers/infiniband/hw/mlx4/mlx4_ib.h
@@ -176,6 +176,10 @@ enum mlx4_ib_qp_type {
 	MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER | \
 	MLX4_IB_QPT_TUN_SMI | MLX4_IB_QPT_TUN_GSI)
 
+enum {
+	MLX4_NUM_TUNNEL_BUFS		= 256,
+};
+
 struct mlx4_ib_tunnel_header {
 	struct mlx4_av av;
 	__be32 remote_qpn;
@@ -263,6 +267,15 @@ struct mlx4_ib_ah {
 	union mlx4_ext_av       av;
 };
 
+struct mlx4_ib_demux_work {
+	struct work_struct	work;
+	struct mlx4_ib_dev     *dev;
+	int			slave;
+	int			do_init;
+	u8			port;
+
+};
+
 struct mlx4_ib_tun_tx_buf {
 	struct mlx4_ib_buf buf;
 	struct ib_ah *ah;
@@ -278,9 +291,17 @@ struct mlx4_ib_demux_pv_qp {
 	unsigned tx_ix_tail;
 };
 
+enum mlx4_ib_demux_pv_state {
+	DEMUX_PV_STATE_DOWN,
+	DEMUX_PV_STATE_STARTING,
+	DEMUX_PV_STATE_ACTIVE,
+	DEMUX_PV_STATE_DOWNING,
+};
+
 struct mlx4_ib_demux_pv_ctx {
 	int port;
 	int slave;
+	enum mlx4_ib_demux_pv_state state;
 	int has_smi;
 	struct ib_device *ib_dev;
 	struct ib_cq *cq;
@@ -319,6 +340,13 @@ struct mlx4_ib_iboe {
 	union ib_gid		gid_table[MLX4_MAX_PORTS][128];
 };
 
+struct pkey_mgt {
+	u8			virt2phys_pkey[MLX4_MFUNC_MAX][MLX4_MAX_PORTS][MLX4_MAX_PORT_PKEYS];
+	u16			phys_pkey_cache[MLX4_MAX_PORTS][MLX4_MAX_PORT_PKEYS];
+	struct list_head	pkey_port_list[MLX4_MFUNC_MAX];
+	struct kobject	       *device_parent[MLX4_MFUNC_MAX];
+};
+
 struct mlx4_ib_dev {
 	struct ib_device	ib_dev;
 	struct mlx4_dev	       *dev;
@@ -340,6 +368,7 @@ struct mlx4_ib_dev {
 	int			counters[MLX4_MAX_PORTS];
 	int		       *eq_table;
 	int			eq_added;
+	struct pkey_mgt		pkeys;
 };
 
 struct ib_event_work {
@@ -424,6 +453,9 @@ static inline struct mlx4_ib_ah *to_mah(struct ib_ah *ibah)
 	return container_of(ibah, struct mlx4_ib_ah, ibah);
 }
 
+int mlx4_ib_init_sriov(struct mlx4_ib_dev *dev);
+void mlx4_ib_close_sriov(struct mlx4_ib_dev *dev);
+
 int mlx4_ib_db_map_user(struct mlx4_ib_ucontext *context, unsigned long virt,
 			struct mlx4_db *db);
 void mlx4_ib_db_unmap_user(struct mlx4_ib_ucontext *context, struct mlx4_db *db);
@@ -515,4 +547,6 @@ int mlx4_ib_add_mc(struct mlx4_ib_dev *mdev, struct mlx4_ib_qp *mqp,
 void mlx4_ib_dispatch_event(struct mlx4_ib_dev *dev, u8 port_num,
 			    enum ib_event_type type);
 
+void mlx4_ib_tunnels_update_work(struct work_struct *work);
+
 #endif /* MLX4_IB_H */
diff --git a/drivers/net/ethernet/mellanox/mlx4/cmd.c b/drivers/net/ethernet/mellanox/mlx4/cmd.c
index c8fef4353021..cb9bebe28276 100644
--- a/drivers/net/ethernet/mellanox/mlx4/cmd.c
+++ b/drivers/net/ethernet/mellanox/mlx4/cmd.c
@@ -1340,6 +1340,8 @@ static void mlx4_master_do_cmd(struct mlx4_dev *dev, int slave, u8 cmd,
 		if (MLX4_COMM_CMD_FLR == slave_state[slave].last_cmd)
 			goto inform_slave_state;
 
+		mlx4_dispatch_event(dev, MLX4_DEV_EVENT_SLAVE_SHUTDOWN, slave);
+
 		/* write the version in the event field */
 		reply |= mlx4_comm_get_version();
 
@@ -1376,6 +1378,7 @@ static void mlx4_master_do_cmd(struct mlx4_dev *dev, int slave, u8 cmd,
 			goto reset_slave;
 		slave_state[slave].vhcr_dma |= param;
 		slave_state[slave].active = true;
+		mlx4_dispatch_event(dev, MLX4_DEV_EVENT_SLAVE_INIT, slave);
 		break;
 	case MLX4_COMM_CMD_VHCR_POST:
 		if ((slave_state[slave].last_cmd != MLX4_COMM_CMD_VHCR_EN) &&
diff --git a/include/linux/mlx4/device.h b/include/linux/mlx4/device.h
index d5c82b7216de..b6b8d341b6c8 100644
--- a/include/linux/mlx4/device.h
+++ b/include/linux/mlx4/device.h
@@ -54,7 +54,8 @@ enum {
 };
 
 enum {
-	MLX4_MAX_PORTS		= 2
+	MLX4_MAX_PORTS		= 2,
+	MLX4_MAX_PORT_PKEYS	= 128
 };
 
 /* base qkey for use in sriov tunnel-qp/proxy-qp communication.
diff --git a/include/linux/mlx4/driver.h b/include/linux/mlx4/driver.h
index d813704b963b..c257e1b211be 100644
--- a/include/linux/mlx4/driver.h
+++ b/include/linux/mlx4/driver.h
@@ -45,6 +45,8 @@ enum mlx4_dev_event {
 	MLX4_DEV_EVENT_PORT_DOWN,
 	MLX4_DEV_EVENT_PORT_REINIT,
 	MLX4_DEV_EVENT_PORT_MGMT_CHANGE,
+	MLX4_DEV_EVENT_SLAVE_INIT,
+	MLX4_DEV_EVENT_SLAVE_SHUTDOWN,
 };
 
 struct mlx4_interface {