[PATCH for-next 00/10] IB/mlx5: Use wr.ud.pkey_index in GSI QP

[PATCH for-next 00/10] IB/mlx5: Use wr.ud.pkey_index in GSI QP

[Haggai Eran]

The check of BTH P_Key for RDMA_CM requests was introduced in patch
4c21b5bcef73 ("IB/cma: Add net_dev and private data checks to RDMA CM").
A side effect of that change was the unexpected behavior of the mlx5 driver
in handling the pkey_index in ib_send_wr.ud in GSI packets.  The pkey_index
in ib_wc was not set correctly [1], and in addition, we found out that the
mlx5 driver did not respect the pkey_index in ib_send_wr.ud for GSI
packets. Since the pkey_index in a Work Request is not explicitly required
by the IBA specifications, the Connect-IB and ConnectX-4 adapter cards do
not support it.

This patch fixes the issue inside the mlx5_ib module. It emulates a single
GSI QP by creating multiple QPs underneath the surface. An ordinary GSI QP
is created for receiving MADs destined at QP 1, and multiple transmission
QPs are created for sending MADs. These QPs behave like UD QPs, although
they use QP 1 as their source QP number. The QPs are created on-demand, one
for each non-zero P_Key in the P_Key table.

The driver handles reordering of the completions to make sure they are
reported in the same order their matching Work Requests were posted.

A spinlock in the emulated GSI QP struct protects concurrent access to the
transmission QPs from post send operations and from their creation
following a P_Key table change. In addition, it protects the Work Request
list that maintains the completion order.

The patchset is divided as follows. The first patch adds the infrastructure
required for creating additional QPs that send as QP 1. The following patch
adds some helpful debugging prints to the QP modification code. Patch 3
adds a skeleton code for an emulated GSI QP. The next two patches create
the multiple transmission QPs. Patches 6-7 handle the reordering of the
completions. Patches 8-9 enable the use of the created QPs and remove the
send queue buffer from the receive GSI QP. Finally patch 10 reverts the
change in "IB/cma: Use inner P_Key to determine netdev".

These patches apply against kernel v4.5-rc6.

[1] mlx5: Fix incorrect wc pkey_index assignment for GSI messages
http://www.spinics.net/lists/linux-rdma/msg28374.html

Haggai Eran (10):
  IB/mlx5: Add support for setting source QP number
  IB/mlx5: Modify QP debugging prints
  IB/mlx5: Add GSI QP wrapper
  IB/mlx5: Create multiple transmission GSI QPs
  IB/mlx5: Create GSI transmission QPs when P_Key table is changed
  IB/mlx5: Generate completions in software
  IB/mlx5: Reorder GSI completions
  IB/mlx5: Pick the right GSI transmission QP for sending
  IB/mlx5: Eliminate GSI RX QP's send buffers
  IB/cma: Print warning on different inner and header P_Keys

 drivers/infiniband/core/cma.c        |   8 +
 drivers/infiniband/hw/mlx5/Makefile  |   2 +-
 drivers/infiniband/hw/mlx5/cq.c      |  83 +++++-
 drivers/infiniband/hw/mlx5/gsi.c     | 548 +++++++++++++++++++++++++++++++++++
 drivers/infiniband/hw/mlx5/main.c    |  30 ++
 drivers/infiniband/hw/mlx5/mlx5_ib.h |  57 +++-
 drivers/infiniband/hw/mlx5/qp.c      |  95 ++++--
 include/linux/mlx5/mlx5_ifc.h        |   3 +-
 include/linux/mlx5/qp.h              |   3 +-
 9 files changed, 802 insertions(+), 27 deletions(-)
 create mode 100644 drivers/infiniband/hw/mlx5/gsi.c

-- 
1.7.11.2

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH for-next 01/10] IB/mlx5: Add support for setting source QP number

[Haggai Eran]

In order to create multiple GSI QPs, we need to set the source QP number to
one on all these QPs. Add the necessary definitions and infrastructure to
do that.

Reviewed-by: Leon Romanovsky <leonro-VPRAkNaXOzVWk0Htik3J/w@public.gmane.org>
Signed-off-by: Haggai Eran <haggaie-VPRAkNaXOzVWk0Htik3J/w@public.gmane.org>
---
 drivers/infiniband/hw/mlx5/mlx5_ib.h | 14 ++++++++++++++
 drivers/infiniband/hw/mlx5/qp.c      | 18 +++++++++++++++++-
 include/linux/mlx5/mlx5_ifc.h        |  3 ++-
 include/linux/mlx5/qp.h              |  3 ++-
 4 files changed, 35 insertions(+), 3 deletions(-)

diff --git a/drivers/infiniband/hw/mlx5/mlx5_ib.h b/drivers/infiniband/hw/mlx5/mlx5_ib.h
index d2b9737baa36..8dc6cd818baf 100644
--- a/drivers/infiniband/hw/mlx5/mlx5_ib.h
+++ b/drivers/infiniband/hw/mlx5/mlx5_ib.h
@@ -165,6 +165,18 @@ struct mlx5_ib_flow_db {
 #define MLX5_IB_QPT_REG_UMR	IB_QPT_RESERVED1
 #define MLX5_IB_WR_UMR		IB_WR_RESERVED1
 
+/* Private QP creation flags to be passed in ib_qp_init_attr.create_flags.
+ *
+ * These flags are intended for internal use by the mlx5_ib driver, and they
+ * rely on the range reserved for that use in the ib_qp_create_flags enum.
+ */
+
+/* Create a UD QP whose source QP number is 1 */
+static inline enum ib_qp_create_flags mlx5_ib_create_qp_sqpn_qp1(void)
+{
+	return IB_QP_CREATE_RESERVED_START;
+}
+
 struct wr_list {
 	u16	opcode;
 	u16	next;
@@ -330,6 +342,8 @@ enum mlx5_ib_qp_flags {
 	MLX5_IB_QP_CROSS_CHANNEL		= 1 << 2,
 	MLX5_IB_QP_MANAGED_SEND			= 1 << 3,
 	MLX5_IB_QP_MANAGED_RECV			= 1 << 4,
+	/* QP uses 1 as its source QP number */
+	MLX5_IB_QP_SQPN_QP1			= 1 << 5,
 };
 
 struct mlx5_umr_wr {
diff --git a/drivers/infiniband/hw/mlx5/qp.c b/drivers/infiniband/hw/mlx5/qp.c
index 34cb8e87c7b8..7cf789d1aa29 100644
--- a/drivers/infiniband/hw/mlx5/qp.c
+++ b/drivers/infiniband/hw/mlx5/qp.c
@@ -783,7 +783,9 @@ static int create_kernel_qp(struct mlx5_ib_dev *dev,
 	int err;
 
 	uuari = &dev->mdev->priv.uuari;
-	if (init_attr->create_flags & ~(IB_QP_CREATE_SIGNATURE_EN | IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK))
+	if (init_attr->create_flags & ~(IB_QP_CREATE_SIGNATURE_EN |
+					IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK |
+					mlx5_ib_create_qp_sqpn_qp1()))
 		return -EINVAL;
 
 	if (init_attr->qp_type == MLX5_IB_QPT_REG_UMR)
@@ -828,6 +830,11 @@ static int create_kernel_qp(struct mlx5_ib_dev *dev,
 	(*in)->ctx.params1 |= cpu_to_be32(1 << 11);
 	(*in)->ctx.sq_crq_size |= cpu_to_be16(1 << 4);
 
+	if (init_attr->create_flags & mlx5_ib_create_qp_sqpn_qp1()) {
+		(*in)->ctx.deth_sqpn = cpu_to_be32(1);
+		qp->flags |= MLX5_IB_QP_SQPN_QP1;
+	}
+
 	mlx5_fill_page_array(&qp->buf, (*in)->pas);
 
 	err = mlx5_db_alloc(dev->mdev, &qp->db);
@@ -1271,6 +1278,11 @@ static int create_qp_common(struct mlx5_ib_dev *dev, struct ib_pd *pd,
 					    ucmd.sq_wqe_count, max_wqes);
 				return -EINVAL;
 			}
+			if (init_attr->create_flags &
+			    mlx5_ib_create_qp_sqpn_qp1()) {
+				mlx5_ib_dbg(dev, "user-space is not allowed to create UD QPs spoofing as QP1\n");
+				return -EINVAL;
+			}
 			err = create_user_qp(dev, pd, qp, udata, init_attr, &in,
 					     &resp, &inlen, base);
 			if (err)
@@ -2284,6 +2296,8 @@ static int __mlx5_ib_modify_qp(struct ib_qp *ibqp,
 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
 		context->sq_crq_size |= cpu_to_be16(1 << 4);
 
+	if (qp->flags & MLX5_IB_QP_SQPN_QP1)
+		context->deth_sqpn = cpu_to_be32(1);
 
 	mlx5_cur = to_mlx5_state(cur_state);
 	mlx5_new = to_mlx5_state(new_state);
@@ -3874,6 +3888,8 @@ int mlx5_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
 		qp_init_attr->create_flags |= IB_QP_CREATE_MANAGED_SEND;
 	if (qp->flags & MLX5_IB_QP_MANAGED_RECV)
 		qp_init_attr->create_flags |= IB_QP_CREATE_MANAGED_RECV;
+	if (qp->flags & MLX5_IB_QP_SQPN_QP1)
+		qp_init_attr->create_flags |= mlx5_ib_create_qp_sqpn_qp1();
 
 	qp_init_attr->sq_sig_type = qp->sq_signal_bits & MLX5_WQE_CTRL_CQ_UPDATE ?
 		IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
diff --git a/include/linux/mlx5/mlx5_ifc.h b/include/linux/mlx5/mlx5_ifc.h
index 51f1e540fc2b..2b35ee445cc3 100644
--- a/include/linux/mlx5/mlx5_ifc.h
+++ b/include/linux/mlx5/mlx5_ifc.h
@@ -770,7 +770,8 @@ struct mlx5_ifc_cmd_hca_cap_bits {
 	u8         reserved_at_22e[0x7];
 	u8         qkv[0x1];
 	u8         pkv[0x1];
-	u8         reserved_at_237[0x4];
+	u8         set_deth_sqpn[0x1];
+	u8         reserved_at_239[0x3];
 	u8         xrc[0x1];
 	u8         ud[0x1];
 	u8         uc[0x1];
diff --git a/include/linux/mlx5/qp.h b/include/linux/mlx5/qp.h
index 5b8c89ffaa58..e5bbcf06de95 100644
--- a/include/linux/mlx5/qp.h
+++ b/include/linux/mlx5/qp.h
@@ -499,7 +499,8 @@ struct mlx5_qp_context {
 	u8			reserved2[4];
 	__be32			next_send_psn;
 	__be32			cqn_send;
-	u8			reserved3[8];
+	__be32			deth_sqpn;
+	u8			reserved3[4];
 	__be32			last_acked_psn;
 	__be32			ssn;
 	__be32			params2;
-- 
1.7.11.2

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH for-next 02/10] IB/mlx5: Modify QP debugging prints

[Haggai Eran]

Add debugging prints to the modify QP verb to help understand the cause a
returned error.

Reviewed-by: Leon Romanovsky <leonro-VPRAkNaXOzVWk0Htik3J/w@public.gmane.org>
Signed-off-by: Haggai Eran <haggaie-VPRAkNaXOzVWk0Htik3J/w@public.gmane.org>
---
 drivers/infiniband/hw/mlx5/qp.c | 29 +++++++++++++++++++++++------
 1 file changed, 23 insertions(+), 6 deletions(-)

diff --git a/drivers/infiniband/hw/mlx5/qp.c b/drivers/infiniband/hw/mlx5/qp.c
index 7cf789d1aa29..784ddf77fca2 100644
--- a/drivers/infiniband/hw/mlx5/qp.c
+++ b/drivers/infiniband/hw/mlx5/qp.c
@@ -2173,8 +2173,10 @@ static int __mlx5_ib_modify_qp(struct ib_qp *ibqp,
 
 	context = &in->ctx;
 	err = to_mlx5_st(ibqp->qp_type);
-	if (err < 0)
+	if (err < 0) {
+		mlx5_ib_dbg(dev, "unsupported qp type %d\n", ibqp->qp_type);
 		goto out;
+	}
 
 	context->flags = cpu_to_be32(err << 16);
 
@@ -2394,30 +2396,45 @@ int mlx5_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 
 	if (ibqp->qp_type != MLX5_IB_QPT_REG_UMR &&
 	    !ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, attr_mask,
-				ll))
+				ll)) {
+		mlx5_ib_dbg(dev, "invalid QP state transition from %d to %d, qp_type %d, attr_mask 0x%x\n",
+			    cur_state, new_state, ibqp->qp_type, attr_mask);
 		goto out;
+	}
 
 	if ((attr_mask & IB_QP_PORT) &&
 	    (attr->port_num == 0 ||
-	     attr->port_num > MLX5_CAP_GEN(dev->mdev, num_ports)))
+	     attr->port_num > MLX5_CAP_GEN(dev->mdev, num_ports))) {
+		mlx5_ib_dbg(dev, "invalid port number %d. number of ports is %d\n",
+			    attr->port_num, dev->num_ports);
 		goto out;
+	}
 
 	if (attr_mask & IB_QP_PKEY_INDEX) {
 		port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
 		if (attr->pkey_index >=
-		    dev->mdev->port_caps[port - 1].pkey_table_len)
+		    dev->mdev->port_caps[port - 1].pkey_table_len) {
+			mlx5_ib_dbg(dev, "invalid pkey index %d\n",
+				    attr->pkey_index);
 			goto out;
+		}
 	}
 
 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
 	    attr->max_rd_atomic >
-	    (1 << MLX5_CAP_GEN(dev->mdev, log_max_ra_res_qp)))
+	    (1 << MLX5_CAP_GEN(dev->mdev, log_max_ra_res_qp))) {
+		mlx5_ib_dbg(dev, "invalid max_rd_atomic value %d\n",
+			    attr->max_rd_atomic);
 		goto out;
+	}
 
 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
 	    attr->max_dest_rd_atomic >
-	    (1 << MLX5_CAP_GEN(dev->mdev, log_max_ra_req_qp)))
+	    (1 << MLX5_CAP_GEN(dev->mdev, log_max_ra_req_qp))) {
+		mlx5_ib_dbg(dev, "invalid max_dest_rd_atomic value %d\n",
+			    attr->max_dest_rd_atomic);
 		goto out;
+	}
 
 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
 		err = 0;
-- 
1.7.11.2

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH for-next 03/10] IB/mlx5: Add GSI QP wrapper

[Haggai Eran]

mlx5 creates special GSI QPs that has limited ability to control the P_Key
of transmitted packets. The sent P_Key is taken from the QP object,
similarly to what happens with regular UD QPs.

Create a software wrapper around GSI QPs that with the following patches
will be able to emulate the functionality of a GSI QP including control of
the P_Key per work request.

Reviewed-by: Leon Romanovsky <leonro-VPRAkNaXOzVWk0Htik3J/w@public.gmane.org>
Signed-off-by: Haggai Eran <haggaie-VPRAkNaXOzVWk0Htik3J/w@public.gmane.org>
---
 drivers/infiniband/hw/mlx5/Makefile  |   2 +-
 drivers/infiniband/hw/mlx5/gsi.c     | 179 +++++++++++++++++++++++++++++++++++
 drivers/infiniband/hw/mlx5/main.c    |   2 +
 drivers/infiniband/hw/mlx5/mlx5_ib.h |  30 +++++-
 drivers/infiniband/hw/mlx5/qp.c      |  50 +++++++---
 5 files changed, 249 insertions(+), 14 deletions(-)
 create mode 100644 drivers/infiniband/hw/mlx5/gsi.c

diff --git a/drivers/infiniband/hw/mlx5/Makefile b/drivers/infiniband/hw/mlx5/Makefile
index 27a70159e2ea..4e851889355a 100644
--- a/drivers/infiniband/hw/mlx5/Makefile
+++ b/drivers/infiniband/hw/mlx5/Makefile
@@ -1,4 +1,4 @@
 obj-$(CONFIG_MLX5_INFINIBAND)	+= mlx5_ib.o
 
-mlx5_ib-y :=	main.o cq.o doorbell.o qp.o mem.o srq.o mr.o ah.o mad.o
+mlx5_ib-y :=	main.o cq.o doorbell.o qp.o mem.o srq.o mr.o ah.o mad.o gsi.o
 mlx5_ib-$(CONFIG_INFINIBAND_ON_DEMAND_PAGING) += odp.o
diff --git a/drivers/infiniband/hw/mlx5/gsi.c b/drivers/infiniband/hw/mlx5/gsi.c
new file mode 100644
index 000000000000..711655400ca5
--- /dev/null
+++ b/drivers/infiniband/hw/mlx5/gsi.c
@@ -0,0 +1,179 @@
+/*
+ * Copyright (c) 2016, Mellanox Technologies. 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 "mlx5_ib.h"
+
+struct mlx5_ib_gsi_qp {
+	struct ib_qp ibqp;
+	struct ib_qp *rx_qp;
+	u8 port_num;
+	struct ib_qp_cap cap;
+	enum ib_sig_type sq_sig_type;
+	/* Serialize qp state modifications */
+	struct mutex mutex;
+};
+
+static struct mlx5_ib_gsi_qp *gsi_qp(struct ib_qp *qp)
+{
+	return container_of(qp, struct mlx5_ib_gsi_qp, ibqp);
+}
+
+struct ib_qp *mlx5_ib_gsi_create_qp(struct ib_pd *pd,
+				    struct ib_qp_init_attr *init_attr)
+{
+	struct mlx5_ib_dev *dev = to_mdev(pd->device);
+	struct mlx5_ib_gsi_qp *gsi;
+	struct ib_qp_init_attr hw_init_attr = *init_attr;
+	const u8 port_num = init_attr->port_num;
+	int ret;
+
+	mlx5_ib_dbg(dev, "creating GSI QP\n");
+
+	if (port_num > ARRAY_SIZE(dev->devr.ports) || port_num < 1) {
+		mlx5_ib_warn(dev,
+			     "invalid port number %d during GSI QP creation\n",
+			     port_num);
+		return ERR_PTR(-EINVAL);
+	}
+
+	gsi = kzalloc(sizeof(*gsi), GFP_KERNEL);
+	if (!gsi)
+		return ERR_PTR(-ENOMEM);
+
+	mutex_init(&gsi->mutex);
+
+	mutex_lock(&dev->devr.mutex);
+
+	if (dev->devr.ports[port_num - 1].gsi) {
+		mlx5_ib_warn(dev, "GSI QP already exists on port %d\n",
+			     port_num);
+		ret = -EBUSY;
+		goto err_free;
+	}
+
+	gsi->cap = init_attr->cap;
+	gsi->sq_sig_type = init_attr->sq_sig_type;
+	gsi->ibqp.qp_num = 1;
+	gsi->port_num = port_num;
+
+	hw_init_attr.qp_type = MLX5_IB_QPT_HW_GSI;
+	gsi->rx_qp = ib_create_qp(pd, &hw_init_attr);
+	if (IS_ERR(gsi->rx_qp)) {
+		mlx5_ib_warn(dev, "unable to create hardware GSI QP. error %ld\n",
+			     PTR_ERR(gsi->rx_qp));
+		ret = PTR_ERR(gsi->rx_qp);
+		goto err_free;
+	}
+
+	dev->devr.ports[init_attr->port_num - 1].gsi = gsi;
+
+	mutex_unlock(&dev->devr.mutex);
+
+	return &gsi->ibqp;
+
+err_free:
+	mutex_unlock(&dev->devr.mutex);
+	kfree(gsi);
+	return ERR_PTR(ret);
+}
+
+int mlx5_ib_gsi_destroy_qp(struct ib_qp *qp)
+{
+	struct mlx5_ib_dev *dev = to_mdev(qp->device);
+	struct mlx5_ib_gsi_qp *gsi = gsi_qp(qp);
+	const int port_num = gsi->port_num;
+	int ret;
+
+	mlx5_ib_dbg(dev, "destroying GSI QP\n");
+
+	mutex_lock(&dev->devr.mutex);
+	ret = ib_destroy_qp(gsi->rx_qp);
+	if (ret) {
+		mlx5_ib_warn(dev, "unable to destroy hardware GSI QP. error %d\n",
+			     ret);
+		mutex_unlock(&dev->devr.mutex);
+		return ret;
+	}
+	dev->devr.ports[port_num - 1].gsi = NULL;
+	mutex_unlock(&dev->devr.mutex);
+
+	kfree(gsi);
+
+	return 0;
+}
+
+int mlx5_ib_gsi_modify_qp(struct ib_qp *qp, struct ib_qp_attr *attr,
+			  int attr_mask)
+{
+	struct mlx5_ib_dev *dev = to_mdev(qp->device);
+	struct mlx5_ib_gsi_qp *gsi = gsi_qp(qp);
+	int ret;
+
+	mlx5_ib_dbg(dev, "modifying GSI QP to state %d\n", attr->qp_state);
+
+	mutex_lock(&gsi->mutex);
+	ret = ib_modify_qp(gsi->rx_qp, attr, attr_mask);
+	mutex_unlock(&gsi->mutex);
+
+	return ret;
+}
+
+int mlx5_ib_gsi_query_qp(struct ib_qp *qp, struct ib_qp_attr *qp_attr,
+			 int qp_attr_mask,
+			 struct ib_qp_init_attr *qp_init_attr)
+{
+	struct mlx5_ib_gsi_qp *gsi = gsi_qp(qp);
+	int ret;
+
+	mutex_lock(&gsi->mutex);
+	ret = ib_query_qp(gsi->rx_qp, qp_attr, qp_attr_mask, qp_init_attr);
+	qp_init_attr->cap = gsi->cap;
+	mutex_unlock(&gsi->mutex);
+
+	return ret;
+}
+
+int mlx5_ib_gsi_post_send(struct ib_qp *qp, struct ib_send_wr *wr,
+			  struct ib_send_wr **bad_wr)
+{
+	struct mlx5_ib_gsi_qp *gsi = gsi_qp(qp);
+
+	return ib_post_send(gsi->rx_qp, wr, bad_wr);
+}
+
+int mlx5_ib_gsi_post_recv(struct ib_qp *qp, struct ib_recv_wr *wr,
+			  struct ib_recv_wr **bad_wr)
+{
+	struct mlx5_ib_gsi_qp *gsi = gsi_qp(qp);
+
+	return ib_post_recv(gsi->rx_qp, wr, bad_wr);
+}
diff --git a/drivers/infiniband/hw/mlx5/main.c b/drivers/infiniband/hw/mlx5/main.c
index 03c418ccbc98..a2a16db62b5e 100644
--- a/drivers/infiniband/hw/mlx5/main.c
+++ b/drivers/infiniband/hw/mlx5/main.c
@@ -1965,6 +1965,8 @@ static int create_dev_resources(struct mlx5_ib_resources *devr)
 
 	dev = container_of(devr, struct mlx5_ib_dev, devr);
 
+	mutex_init(&devr->mutex);
+
 	devr->p0 = mlx5_ib_alloc_pd(&dev->ib_dev, NULL, NULL);
 	if (IS_ERR(devr->p0)) {
 		ret = PTR_ERR(devr->p0);
diff --git a/drivers/infiniband/hw/mlx5/mlx5_ib.h b/drivers/infiniband/hw/mlx5/mlx5_ib.h
index 8dc6cd818baf..81af9842ab29 100644
--- a/drivers/infiniband/hw/mlx5/mlx5_ib.h
+++ b/drivers/infiniband/hw/mlx5/mlx5_ib.h
@@ -163,6 +163,11 @@ struct mlx5_ib_flow_db {
 #define MLX5_IB_SEND_UMR_FAIL_IF_FREE (IB_SEND_RESERVED_START << 1)
 #define MLX5_IB_SEND_UMR_UPDATE_MTT (IB_SEND_RESERVED_START << 2)
 #define MLX5_IB_QPT_REG_UMR	IB_QPT_RESERVED1
+/*
+ * IB_QPT_GSI creates the software wrapper around GSI, and MLX5_IB_QPT_HW_GSI
+ * creates the actual hardware QP.
+ */
+#define MLX5_IB_QPT_HW_GSI	IB_QPT_RESERVED2
 #define MLX5_IB_WR_UMR		IB_WR_RESERVED1
 
 /* Private QP creation flags to be passed in ib_qp_init_attr.create_flags.
@@ -501,6 +506,12 @@ struct mlx5_mr_cache {
 	unsigned long		last_add;
 };
 
+struct mlx5_ib_gsi_qp;
+
+struct mlx5_ib_port_resources {
+	struct mlx5_ib_gsi_qp *gsi;
+};
+
 struct mlx5_ib_resources {
 	struct ib_cq	*c0;
 	struct ib_xrcd	*x0;
@@ -508,6 +519,9 @@ struct mlx5_ib_resources {
 	struct ib_pd	*p0;
 	struct ib_srq	*s0;
 	struct ib_srq	*s1;
+	struct mlx5_ib_port_resources ports[2];
+	/* Protects changes to the port resources */
+	struct mutex	mutex;
 };
 
 struct mlx5_roce {
@@ -753,6 +767,20 @@ static inline void mlx5_ib_qp_enable_pagefaults(struct mlx5_ib_qp *qp)  {}
 __be16 mlx5_get_roce_udp_sport(struct mlx5_ib_dev *dev, u8 port_num,
 			       int index);
 
+/* GSI QP helper functions */
+struct ib_qp *mlx5_ib_gsi_create_qp(struct ib_pd *pd,
+				    struct ib_qp_init_attr *init_attr);
+int mlx5_ib_gsi_destroy_qp(struct ib_qp *qp);
+int mlx5_ib_gsi_modify_qp(struct ib_qp *qp, struct ib_qp_attr *attr,
+			  int attr_mask);
+int mlx5_ib_gsi_query_qp(struct ib_qp *qp, struct ib_qp_attr *qp_attr,
+			 int qp_attr_mask,
+			 struct ib_qp_init_attr *qp_init_attr);
+int mlx5_ib_gsi_post_send(struct ib_qp *qp, struct ib_send_wr *wr,
+			  struct ib_send_wr **bad_wr);
+int mlx5_ib_gsi_post_recv(struct ib_qp *qp, struct ib_recv_wr *wr,
+			  struct ib_recv_wr **bad_wr);
+
 static inline void init_query_mad(struct ib_smp *mad)
 {
 	mad->base_version  = 1;
@@ -772,7 +800,7 @@ static inline u8 convert_access(int acc)
 
 static inline int is_qp1(enum ib_qp_type qp_type)
 {
-	return qp_type == IB_QPT_GSI;
+	return qp_type == MLX5_IB_QPT_HW_GSI;
 }
 
 #define MLX5_MAX_UMR_SHIFT 16
diff --git a/drivers/infiniband/hw/mlx5/qp.c b/drivers/infiniband/hw/mlx5/qp.c
index 784ddf77fca2..c1196cc4a5e8 100644
--- a/drivers/infiniband/hw/mlx5/qp.c
+++ b/drivers/infiniband/hw/mlx5/qp.c
@@ -288,7 +288,7 @@ static int sq_overhead(enum ib_qp_type qp_type)
 
 	case IB_QPT_UD:
 	case IB_QPT_SMI:
-	case IB_QPT_GSI:
+	case MLX5_IB_QPT_HW_GSI:
 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
 			sizeof(struct mlx5_wqe_datagram_seg);
 		break;
@@ -590,7 +590,7 @@ static int to_mlx5_st(enum ib_qp_type type)
 	case IB_QPT_XRC_INI:
 	case IB_QPT_XRC_TGT:		return MLX5_QP_ST_XRC;
 	case IB_QPT_SMI:		return MLX5_QP_ST_QP0;
-	case IB_QPT_GSI:		return MLX5_QP_ST_QP1;
+	case MLX5_IB_QPT_HW_GSI:	return MLX5_QP_ST_QP1;
 	case IB_QPT_RAW_IPV6:		return MLX5_QP_ST_RAW_IPV6;
 	case IB_QPT_RAW_PACKET:
 	case IB_QPT_RAW_ETHERTYPE:	return MLX5_QP_ST_RAW_ETHERTYPE;
@@ -1506,7 +1506,7 @@ static void get_cqs(struct mlx5_ib_qp *qp,
 		break;
 
 	case IB_QPT_SMI:
-	case IB_QPT_GSI:
+	case MLX5_IB_QPT_HW_GSI:
 	case IB_QPT_RC:
 	case IB_QPT_UC:
 	case IB_QPT_UD:
@@ -1669,7 +1669,7 @@ struct ib_qp *mlx5_ib_create_qp(struct ib_pd *pd,
 	case IB_QPT_UC:
 	case IB_QPT_UD:
 	case IB_QPT_SMI:
-	case IB_QPT_GSI:
+	case MLX5_IB_QPT_HW_GSI:
 	case MLX5_IB_QPT_REG_UMR:
 		qp = kzalloc(sizeof(*qp), GFP_KERNEL);
 		if (!qp)
@@ -1698,6 +1698,9 @@ struct ib_qp *mlx5_ib_create_qp(struct ib_pd *pd,
 
 		break;
 
+	case IB_QPT_GSI:
+		return mlx5_ib_gsi_create_qp(pd, init_attr);
+
 	case IB_QPT_RAW_IPV6:
 	case IB_QPT_RAW_ETHERTYPE:
 	case IB_QPT_MAX:
@@ -1716,6 +1719,9 @@ int mlx5_ib_destroy_qp(struct ib_qp *qp)
 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
 	struct mlx5_ib_qp *mqp = to_mqp(qp);
 
+	if (unlikely(qp->qp_type == IB_QPT_GSI))
+		return mlx5_ib_gsi_destroy_qp(qp);
+
 	destroy_qp_common(dev, mqp);
 
 	kfree(mqp);
@@ -2196,7 +2202,7 @@ static int __mlx5_ib_modify_qp(struct ib_qp *ibqp,
 		}
 	}
 
-	if (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI) {
+	if (is_sqp(ibqp->qp_type)) {
 		context->mtu_msgmax = (IB_MTU_256 << 5) | 8;
 	} else if (ibqp->qp_type == IB_QPT_UD ||
 		   ibqp->qp_type == MLX5_IB_QPT_REG_UMR) {
@@ -2379,11 +2385,18 @@ int mlx5_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 {
 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
+	enum ib_qp_type qp_type;
 	enum ib_qp_state cur_state, new_state;
 	int err = -EINVAL;
 	int port;
 	enum rdma_link_layer ll = IB_LINK_LAYER_UNSPECIFIED;
 
+	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
+		return mlx5_ib_gsi_modify_qp(ibqp, attr, attr_mask);
+
+	qp_type = (unlikely(ibqp->qp_type == MLX5_IB_QPT_HW_GSI)) ?
+		IB_QPT_GSI : ibqp->qp_type;
+
 	mutex_lock(&qp->mutex);
 
 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
@@ -2394,9 +2407,8 @@ int mlx5_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 		ll = dev->ib_dev.get_link_layer(&dev->ib_dev, port);
 	}
 
-	if (ibqp->qp_type != MLX5_IB_QPT_REG_UMR &&
-	    !ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, attr_mask,
-				ll)) {
+	if (qp_type != MLX5_IB_QPT_REG_UMR &&
+	    !ib_modify_qp_is_ok(cur_state, new_state, qp_type, attr_mask, ll)) {
 		mlx5_ib_dbg(dev, "invalid QP state transition from %d to %d, qp_type %d, attr_mask 0x%x\n",
 			    cur_state, new_state, ibqp->qp_type, attr_mask);
 		goto out;
@@ -3227,13 +3239,13 @@ int mlx5_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
 {
 	struct mlx5_wqe_ctrl_seg *ctrl = NULL;  /* compiler warning */
 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
-	struct mlx5_ib_qp *qp = to_mqp(ibqp);
+	struct mlx5_ib_qp *qp;
 	struct mlx5_ib_mr *mr;
 	struct mlx5_wqe_data_seg *dpseg;
 	struct mlx5_wqe_xrc_seg *xrc;
-	struct mlx5_bf *bf = qp->bf;
+	struct mlx5_bf *bf;
 	int uninitialized_var(size);
-	void *qend = qp->sq.qend;
+	void *qend;
 	unsigned long flags;
 	unsigned idx;
 	int err = 0;
@@ -3245,6 +3257,13 @@ int mlx5_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
 	u8 next_fence = 0;
 	u8 fence;
 
+	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
+		return mlx5_ib_gsi_post_send(ibqp, wr, bad_wr);
+
+	qp = to_mqp(ibqp);
+	bf = qp->bf;
+	qend = qp->sq.qend;
+
 	spin_lock_irqsave(&qp->sq.lock, flags);
 
 	for (nreq = 0; wr; nreq++, wr = wr->next) {
@@ -3406,7 +3425,7 @@ int mlx5_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
 
 		case IB_QPT_UD:
 		case IB_QPT_SMI:
-		case IB_QPT_GSI:
+		case MLX5_IB_QPT_HW_GSI:
 			set_datagram_seg(seg, wr);
 			seg += sizeof(struct mlx5_wqe_datagram_seg);
 			size += sizeof(struct mlx5_wqe_datagram_seg) / 16;
@@ -3533,6 +3552,9 @@ int mlx5_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
 	int ind;
 	int i;
 
+	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
+		return mlx5_ib_gsi_post_recv(ibqp, wr, bad_wr);
+
 	spin_lock_irqsave(&qp->rq.lock, flags);
 
 	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
@@ -3853,6 +3875,10 @@ int mlx5_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
 	int err = 0;
 	u8 raw_packet_qp_state;
 
+	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
+		return mlx5_ib_gsi_query_qp(ibqp, qp_attr, qp_attr_mask,
+					    qp_init_attr);
+
 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
 	/*
 	 * Wait for any outstanding page faults, in case the user frees memory
-- 
1.7.11.2

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH for-next 04/10] IB/mlx5: Create multiple transmission GSI QPs

[Haggai Eran]

In order to send GSI MADs on different P_Keys, mlx5 needs different QPs to
be created, each with a different P_Key set when the QP is modified to the
INIT state.

Create QPs for each non-zero P_Key in the P_Key table.

Signed-off-by: Haggai Eran <haggaie-VPRAkNaXOzVWk0Htik3J/w@public.gmane.org>
---
 drivers/infiniband/hw/mlx5/gsi.c | 170 ++++++++++++++++++++++++++++++++++++++-
 1 file changed, 167 insertions(+), 3 deletions(-)

diff --git a/drivers/infiniband/hw/mlx5/gsi.c b/drivers/infiniband/hw/mlx5/gsi.c
index 711655400ca5..91bd20eb59b7 100644
--- a/drivers/infiniband/hw/mlx5/gsi.c
+++ b/drivers/infiniband/hw/mlx5/gsi.c
@@ -40,6 +40,12 @@ struct mlx5_ib_gsi_qp {
 	enum ib_sig_type sq_sig_type;
 	/* Serialize qp state modifications */
 	struct mutex mutex;
+	int num_qps;
+	/* Protects access to the tx_qps. Post send operations synchronize
+	 * with tx_qp creation in setup_qp().
+	 */
+	spinlock_t lock;
+	struct ib_qp **tx_qps;
 };
 
 static struct mlx5_ib_gsi_qp *gsi_qp(struct ib_qp *qp)
@@ -47,6 +53,11 @@ static struct mlx5_ib_gsi_qp *gsi_qp(struct ib_qp *qp)
 	return container_of(qp, struct mlx5_ib_gsi_qp, ibqp);
 }
 
+static bool mlx5_ib_deth_sqpn_cap(struct mlx5_ib_dev *dev)
+{
+	return MLX5_CAP_GEN(dev->mdev, set_deth_sqpn);
+}
+
 struct ib_qp *mlx5_ib_gsi_create_qp(struct ib_pd *pd,
 				    struct ib_qp_init_attr *init_attr)
 {
@@ -54,6 +65,8 @@ struct ib_qp *mlx5_ib_gsi_create_qp(struct ib_pd *pd,
 	struct mlx5_ib_gsi_qp *gsi;
 	struct ib_qp_init_attr hw_init_attr = *init_attr;
 	const u8 port_num = init_attr->port_num;
+	const int num_pkeys = pd->device->attrs.max_pkeys;
+	const int num_qps = mlx5_ib_deth_sqpn_cap(dev) ? num_pkeys : 0;
 	int ret;
 
 	mlx5_ib_dbg(dev, "creating GSI QP\n");
@@ -69,6 +82,12 @@ struct ib_qp *mlx5_ib_gsi_create_qp(struct ib_pd *pd,
 	if (!gsi)
 		return ERR_PTR(-ENOMEM);
 
+	gsi->tx_qps = kcalloc(num_qps, sizeof(*gsi->tx_qps), GFP_KERNEL);
+	if (!gsi->tx_qps) {
+		ret = -ENOMEM;
+		goto err_free;
+	}
+
 	mutex_init(&gsi->mutex);
 
 	mutex_lock(&dev->devr.mutex);
@@ -77,8 +96,10 @@ struct ib_qp *mlx5_ib_gsi_create_qp(struct ib_pd *pd,
 		mlx5_ib_warn(dev, "GSI QP already exists on port %d\n",
 			     port_num);
 		ret = -EBUSY;
-		goto err_free;
+		goto err_free_tx;
 	}
+	gsi->num_qps = num_qps;
+	spin_lock_init(&gsi->lock);
 
 	gsi->cap = init_attr->cap;
 	gsi->sq_sig_type = init_attr->sq_sig_type;
@@ -91,7 +112,7 @@ struct ib_qp *mlx5_ib_gsi_create_qp(struct ib_pd *pd,
 		mlx5_ib_warn(dev, "unable to create hardware GSI QP. error %ld\n",
 			     PTR_ERR(gsi->rx_qp));
 		ret = PTR_ERR(gsi->rx_qp);
-		goto err_free;
+		goto err_free_tx;
 	}
 
 	dev->devr.ports[init_attr->port_num - 1].gsi = gsi;
@@ -100,8 +121,10 @@ struct ib_qp *mlx5_ib_gsi_create_qp(struct ib_pd *pd,
 
 	return &gsi->ibqp;
 
-err_free:
+err_free_tx:
 	mutex_unlock(&dev->devr.mutex);
+	kfree(gsi->tx_qps);
+err_free:
 	kfree(gsi);
 	return ERR_PTR(ret);
 }
@@ -111,6 +134,7 @@ int mlx5_ib_gsi_destroy_qp(struct ib_qp *qp)
 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
 	struct mlx5_ib_gsi_qp *gsi = gsi_qp(qp);
 	const int port_num = gsi->port_num;
+	int qp_index;
 	int ret;
 
 	mlx5_ib_dbg(dev, "destroying GSI QP\n");
@@ -125,12 +149,143 @@ int mlx5_ib_gsi_destroy_qp(struct ib_qp *qp)
 	}
 	dev->devr.ports[port_num - 1].gsi = NULL;
 	mutex_unlock(&dev->devr.mutex);
+	gsi->rx_qp = NULL;
+
+	for (qp_index = 0; qp_index < gsi->num_qps; ++qp_index) {
+		if (!gsi->tx_qps[qp_index])
+			continue;
+		WARN_ON_ONCE(ib_destroy_qp(gsi->tx_qps[qp_index]));
+		gsi->tx_qps[qp_index] = NULL;
+	}
 
+	kfree(gsi->tx_qps);
 	kfree(gsi);
 
 	return 0;
 }
 
+static struct ib_qp *create_gsi_ud_qp(struct mlx5_ib_gsi_qp *gsi)
+{
+	struct ib_pd *pd = gsi->rx_qp->pd;
+	struct ib_qp_init_attr init_attr = {
+		.event_handler = gsi->rx_qp->event_handler,
+		.qp_context = gsi->rx_qp->qp_context,
+		.send_cq = gsi->rx_qp->send_cq,
+		.recv_cq = gsi->rx_qp->recv_cq,
+		.cap = {
+			.max_send_wr = gsi->cap.max_send_wr,
+			.max_send_sge = gsi->cap.max_send_sge,
+			.max_inline_data = gsi->cap.max_inline_data,
+		},
+		.sq_sig_type = gsi->sq_sig_type,
+		.qp_type = IB_QPT_UD,
+		.create_flags = mlx5_ib_create_qp_sqpn_qp1(),
+	};
+
+	return ib_create_qp(pd, &init_attr);
+}
+
+static int modify_to_rts(struct mlx5_ib_gsi_qp *gsi, struct ib_qp *qp,
+			 u16 qp_index)
+{
+	struct mlx5_ib_dev *dev = to_mdev(qp->device);
+	struct ib_qp_attr attr;
+	int mask;
+	int ret;
+
+	mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_QKEY | IB_QP_PORT;
+	attr.qp_state = IB_QPS_INIT;
+	attr.pkey_index = qp_index;
+	attr.qkey = IB_QP1_QKEY;
+	attr.port_num = gsi->port_num;
+	ret = ib_modify_qp(qp, &attr, mask);
+	if (ret) {
+		mlx5_ib_err(dev, "could not change QP%d state to INIT: %d\n",
+			    qp->qp_num, ret);
+		return ret;
+	}
+
+	attr.qp_state = IB_QPS_RTR;
+	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
+	if (ret) {
+		mlx5_ib_err(dev, "could not change QP%d state to RTR: %d\n",
+			    qp->qp_num, ret);
+		return ret;
+	}
+
+	attr.qp_state = IB_QPS_RTS;
+	attr.sq_psn = 0;
+	ret = ib_modify_qp(qp, &attr, IB_QP_STATE | IB_QP_SQ_PSN);
+	if (ret) {
+		mlx5_ib_err(dev, "could not change QP%d state to RTS: %d\n",
+			    qp->qp_num, ret);
+		return ret;
+	}
+
+	return 0;
+}
+
+static void setup_qp(struct mlx5_ib_gsi_qp *gsi, u16 qp_index)
+{
+	struct ib_device *device = gsi->rx_qp->device;
+	struct mlx5_ib_dev *dev = to_mdev(device);
+	struct ib_qp *qp;
+	unsigned long flags;
+	u16 pkey;
+	int ret;
+
+	ret = ib_query_pkey(device, gsi->port_num, qp_index, &pkey);
+	if (ret) {
+		mlx5_ib_warn(dev, "unable to read P_Key at port %d, index %d\n",
+			     gsi->port_num, qp_index);
+		return;
+	}
+
+	if (!pkey) {
+		mlx5_ib_dbg(dev, "invalid P_Key at port %d, index %d.  Skipping.\n",
+			    gsi->port_num, qp_index);
+		return;
+	}
+
+	spin_lock_irqsave(&gsi->lock, flags);
+	qp = gsi->tx_qps[qp_index];
+	spin_unlock_irqrestore(&gsi->lock, flags);
+	if (qp) {
+		mlx5_ib_dbg(dev, "already existing GSI TX QP at port %d, index %d. Skipping\n",
+			    gsi->port_num, qp_index);
+		return;
+	}
+
+	qp = create_gsi_ud_qp(gsi);
+	if (IS_ERR(qp)) {
+		mlx5_ib_warn(dev, "unable to create hardware UD QP for GSI: %ld\n",
+			     PTR_ERR(qp));
+		return;
+	}
+
+	ret = modify_to_rts(gsi, qp, qp_index);
+	if (ret)
+		goto err_destroy_qp;
+
+	spin_lock_irqsave(&gsi->lock, flags);
+	WARN_ON_ONCE(gsi->tx_qps[qp_index]);
+	gsi->tx_qps[qp_index] = qp;
+	spin_unlock_irqrestore(&gsi->lock, flags);
+
+	return;
+
+err_destroy_qp:
+	WARN_ON_ONCE(qp);
+}
+
+static void setup_qps(struct mlx5_ib_gsi_qp *gsi)
+{
+	u16 qp_index;
+
+	for (qp_index = 0; qp_index < gsi->num_qps; ++qp_index)
+		setup_qp(gsi, qp_index);
+}
+
 int mlx5_ib_gsi_modify_qp(struct ib_qp *qp, struct ib_qp_attr *attr,
 			  int attr_mask)
 {
@@ -142,6 +297,15 @@ int mlx5_ib_gsi_modify_qp(struct ib_qp *qp, struct ib_qp_attr *attr,
 
 	mutex_lock(&gsi->mutex);
 	ret = ib_modify_qp(gsi->rx_qp, attr, attr_mask);
+	if (ret) {
+		mlx5_ib_warn(dev, "unable to modify GSI rx QP: %d\n", ret);
+		goto unlock;
+	}
+
+	if (to_mqp(gsi->rx_qp)->state == IB_QPS_RTS)
+		setup_qps(gsi);
+
+unlock:
 	mutex_unlock(&gsi->mutex);
 
 	return ret;
-- 
1.7.11.2

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH for-next 05/10] IB/mlx5: Create GSI transmission QPs when P_Key table is changed

[Haggai Eran]

Whenever the P_Key table is changed, we create the required GSI
transmission QPs on-demand.

Reviewed-by: Leon Romanovsky <leonro-VPRAkNaXOzVWk0Htik3J/w@public.gmane.org>
Signed-off-by: Haggai Eran <haggaie-VPRAkNaXOzVWk0Htik3J/w@public.gmane.org>
---
 drivers/infiniband/hw/mlx5/gsi.c     | 10 ++++++++++
 drivers/infiniband/hw/mlx5/main.c    | 28 ++++++++++++++++++++++++++++
 drivers/infiniband/hw/mlx5/mlx5_ib.h |  3 +++
 3 files changed, 41 insertions(+)

diff --git a/drivers/infiniband/hw/mlx5/gsi.c b/drivers/infiniband/hw/mlx5/gsi.c
index 91bd20eb59b7..1648f539c836 100644
--- a/drivers/infiniband/hw/mlx5/gsi.c
+++ b/drivers/infiniband/hw/mlx5/gsi.c
@@ -341,3 +341,13 @@ int mlx5_ib_gsi_post_recv(struct ib_qp *qp, struct ib_recv_wr *wr,
 
 	return ib_post_recv(gsi->rx_qp, wr, bad_wr);
 }
+
+void mlx5_ib_gsi_pkey_change(struct mlx5_ib_gsi_qp *gsi)
+{
+	if (!gsi)
+		return;
+
+	mutex_lock(&gsi->mutex);
+	setup_qps(gsi);
+	mutex_unlock(&gsi->mutex);
+}
diff --git a/drivers/infiniband/hw/mlx5/main.c b/drivers/infiniband/hw/mlx5/main.c
index a2a16db62b5e..3fbc257ba5e2 100644
--- a/drivers/infiniband/hw/mlx5/main.c
+++ b/drivers/infiniband/hw/mlx5/main.c
@@ -1716,6 +1716,17 @@ static struct device_attribute *mlx5_class_attributes[] = {
 	&dev_attr_reg_pages,
 };
 
+static void pkey_change_handler(struct work_struct *work)
+{
+	struct mlx5_ib_port_resources *ports =
+		container_of(work, struct mlx5_ib_port_resources,
+			     pkey_change_work);
+
+	mutex_lock(&ports->devr->mutex);
+	mlx5_ib_gsi_pkey_change(ports->gsi);
+	mutex_unlock(&ports->devr->mutex);
+}
+
 static void mlx5_ib_event(struct mlx5_core_dev *dev, void *context,
 			  enum mlx5_dev_event event, unsigned long param)
 {
@@ -1752,6 +1763,8 @@ static void mlx5_ib_event(struct mlx5_core_dev *dev, void *context,
 	case MLX5_DEV_EVENT_PKEY_CHANGE:
 		ibev.event = IB_EVENT_PKEY_CHANGE;
 		port = (u8)param;
+
+		schedule_work(&ibdev->devr.ports[port - 1].pkey_change_work);
 		break;
 
 	case MLX5_DEV_EVENT_GUID_CHANGE:
@@ -1961,6 +1974,7 @@ static int create_dev_resources(struct mlx5_ib_resources *devr)
 	struct ib_srq_init_attr attr;
 	struct mlx5_ib_dev *dev;
 	struct ib_cq_init_attr cq_attr = {.cqe = 1};
+	int port;
 	int ret = 0;
 
 	dev = container_of(devr, struct mlx5_ib_dev, devr);
@@ -2054,6 +2068,12 @@ static int create_dev_resources(struct mlx5_ib_resources *devr)
 	atomic_inc(&devr->p0->usecnt);
 	atomic_set(&devr->s0->usecnt, 0);
 
+	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port) {
+		INIT_WORK(&devr->ports[port].pkey_change_work,
+			  pkey_change_handler);
+		devr->ports[port].devr = devr;
+	}
+
 	return 0;
 
 error5:
@@ -2072,12 +2092,20 @@ error0:
 
 static void destroy_dev_resources(struct mlx5_ib_resources *devr)
 {
+	struct mlx5_ib_dev *dev =
+		container_of(devr, struct mlx5_ib_dev, devr);
+	int port;
+
 	mlx5_ib_destroy_srq(devr->s1);
 	mlx5_ib_destroy_srq(devr->s0);
 	mlx5_ib_dealloc_xrcd(devr->x0);
 	mlx5_ib_dealloc_xrcd(devr->x1);
 	mlx5_ib_destroy_cq(devr->c0);
 	mlx5_ib_dealloc_pd(devr->p0);
+
+	/* Make sure no change P_Key work items are still executing */
+	for (port = 0; port < dev->num_ports; ++port)
+		cancel_work_sync(&devr->ports[port].pkey_change_work);
 }
 
 static u32 get_core_cap_flags(struct ib_device *ibdev)
diff --git a/drivers/infiniband/hw/mlx5/mlx5_ib.h b/drivers/infiniband/hw/mlx5/mlx5_ib.h
index 81af9842ab29..640840dfc4f4 100644
--- a/drivers/infiniband/hw/mlx5/mlx5_ib.h
+++ b/drivers/infiniband/hw/mlx5/mlx5_ib.h
@@ -509,7 +509,9 @@ struct mlx5_mr_cache {
 struct mlx5_ib_gsi_qp;
 
 struct mlx5_ib_port_resources {
+	struct mlx5_ib_resources *devr;
 	struct mlx5_ib_gsi_qp *gsi;
+	struct work_struct pkey_change_work;
 };
 
 struct mlx5_ib_resources {
@@ -780,6 +782,7 @@ int mlx5_ib_gsi_post_send(struct ib_qp *qp, struct ib_send_wr *wr,
 			  struct ib_send_wr **bad_wr);
 int mlx5_ib_gsi_post_recv(struct ib_qp *qp, struct ib_recv_wr *wr,
 			  struct ib_recv_wr **bad_wr);
+void mlx5_ib_gsi_pkey_change(struct mlx5_ib_gsi_qp *gsi);
 
 static inline void init_query_mad(struct ib_smp *mad)
 {
-- 
1.7.11.2

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH for-next 06/10] IB/mlx5: Generate completions in software

[Haggai Eran]

The GSI QP emulation requires also emulating completions for transmitted
MADs. The CQ on which these completions are generated can also be used by
the hardware, and the MAD layer is free to use any CQ of the device for the
GSI QP.

Add a method for generating software completions to each mlx5 CQ. Software
completions are polled first, and generate calls to the completion handler
callback if necessary.

Reviewed-by: Leon Romanovsky <leonro-VPRAkNaXOzVWk0Htik3J/w@public.gmane.org>
Signed-off-by: Haggai Eran <haggaie-VPRAkNaXOzVWk0Htik3J/w@public.gmane.org>
---
 drivers/infiniband/hw/mlx5/cq.c      | 83 +++++++++++++++++++++++++++++++++---
 drivers/infiniband/hw/mlx5/mlx5_ib.h | 10 +++++
 2 files changed, 88 insertions(+), 5 deletions(-)

diff --git a/drivers/infiniband/hw/mlx5/cq.c b/drivers/infiniband/hw/mlx5/cq.c
index fd1de31e0611..bff5955e9361 100644
--- a/drivers/infiniband/hw/mlx5/cq.c
+++ b/drivers/infiniband/hw/mlx5/cq.c
@@ -565,18 +565,44 @@ repoll:
 	return 0;
 }
 
+static int poll_soft_wc(struct mlx5_ib_cq *cq, int num_entries,
+			struct ib_wc *wc)
+{
+	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
+	struct mlx5_ib_wc *soft_wc, *next;
+	int npolled = 0;
+
+	list_for_each_entry_safe(soft_wc, next, &cq->wc_list, list) {
+		if (npolled >= num_entries)
+			break;
+
+		mlx5_ib_dbg(dev, "polled software generated completion on CQ 0x%x\n",
+			    cq->mcq.cqn);
+
+		wc[npolled++] = soft_wc->wc;
+		list_del(&soft_wc->list);
+		kfree(soft_wc);
+	}
+
+	return npolled;
+}
+
 int mlx5_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
 {
 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
 	struct mlx5_ib_qp *cur_qp = NULL;
 	unsigned long flags;
+	int soft_polled = 0;
 	int npolled;
 	int err = 0;
 
 	spin_lock_irqsave(&cq->lock, flags);
 
-	for (npolled = 0; npolled < num_entries; npolled++) {
-		err = mlx5_poll_one(cq, &cur_qp, wc + npolled);
+	if (unlikely(!list_empty(&cq->wc_list)))
+		soft_polled = poll_soft_wc(cq, num_entries, wc);
+
+	for (npolled = 0; npolled < num_entries - soft_polled; npolled++) {
+		err = mlx5_poll_one(cq, &cur_qp, wc + soft_polled + npolled);
 		if (err)
 			break;
 	}
@@ -587,7 +613,7 @@ int mlx5_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
 	spin_unlock_irqrestore(&cq->lock, flags);
 
 	if (err == 0 || err == -EAGAIN)
-		return npolled;
+		return soft_polled + npolled;
 	else
 		return err;
 }
@@ -595,16 +621,27 @@ int mlx5_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
 int mlx5_ib_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
 {
 	struct mlx5_core_dev *mdev = to_mdev(ibcq->device)->mdev;
+	struct mlx5_ib_cq *cq = to_mcq(ibcq);
 	void __iomem *uar_page = mdev->priv.uuari.uars[0].map;
+	unsigned long irq_flags;
+	int ret = 0;
+
+	spin_lock_irqsave(&cq->lock, irq_flags);
+	if (cq->notify_flags != IB_CQ_NEXT_COMP)
+		cq->notify_flags = flags & IB_CQ_SOLICITED_MASK;
 
-	mlx5_cq_arm(&to_mcq(ibcq)->mcq,
+	if ((flags & IB_CQ_REPORT_MISSED_EVENTS) && !list_empty(&cq->wc_list))
+		ret = 1;
+	spin_unlock_irqrestore(&cq->lock, irq_flags);
+
+	mlx5_cq_arm(&cq->mcq,
 		    (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
 		    MLX5_CQ_DB_REQ_NOT_SOL : MLX5_CQ_DB_REQ_NOT,
 		    uar_page,
 		    MLX5_GET_DOORBELL_LOCK(&mdev->priv.cq_uar_lock),
 		    to_mcq(ibcq)->mcq.cons_index);
 
-	return 0;
+	return ret;
 }
 
 static int alloc_cq_buf(struct mlx5_ib_dev *dev, struct mlx5_ib_cq_buf *buf,
@@ -757,6 +794,14 @@ static void destroy_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq)
 	mlx5_db_free(dev->mdev, &cq->db);
 }
 
+static void notify_soft_wc_handler(struct work_struct *work)
+{
+	struct mlx5_ib_cq *cq = container_of(work, struct mlx5_ib_cq,
+					     notify_work);
+
+	cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
+}
+
 struct ib_cq *mlx5_ib_create_cq(struct ib_device *ibdev,
 				const struct ib_cq_init_attr *attr,
 				struct ib_ucontext *context,
@@ -807,6 +852,8 @@ struct ib_cq *mlx5_ib_create_cq(struct ib_device *ibdev,
 				       &index, &inlen);
 		if (err)
 			goto err_create;
+
+		INIT_WORK(&cq->notify_work, notify_soft_wc_handler);
 	}
 
 	cq->cqe_size = cqe_size;
@@ -832,6 +879,8 @@ struct ib_cq *mlx5_ib_create_cq(struct ib_device *ibdev,
 	cq->mcq.comp  = mlx5_ib_cq_comp;
 	cq->mcq.event = mlx5_ib_cq_event;
 
+	INIT_LIST_HEAD(&cq->wc_list);
+
 	if (context)
 		if (ib_copy_to_udata(udata, &cq->mcq.cqn, sizeof(__u32))) {
 			err = -EFAULT;
@@ -1219,3 +1268,27 @@ int mlx5_ib_get_cqe_size(struct mlx5_ib_dev *dev, struct ib_cq *ibcq)
 	cq = to_mcq(ibcq);
 	return cq->cqe_size;
 }
+
+/* Called from atomic context */
+int mlx5_ib_generate_wc(struct ib_cq *ibcq, struct ib_wc *wc)
+{
+	struct mlx5_ib_wc *soft_wc;
+	struct mlx5_ib_cq *cq = to_mcq(ibcq);
+	unsigned long flags;
+
+	soft_wc = kmalloc(sizeof(*soft_wc), GFP_ATOMIC);
+	if (!soft_wc)
+		return -ENOMEM;
+
+	soft_wc->wc = *wc;
+	spin_lock_irqsave(&cq->lock, flags);
+	list_add_tail(&soft_wc->list, &cq->wc_list);
+	if (cq->notify_flags == IB_CQ_NEXT_COMP ||
+	    wc->status != IB_WC_SUCCESS) {
+		cq->notify_flags = 0;
+		schedule_work(&cq->notify_work);
+	}
+	spin_unlock_irqrestore(&cq->lock, flags);
+
+	return 0;
+}
diff --git a/drivers/infiniband/hw/mlx5/mlx5_ib.h b/drivers/infiniband/hw/mlx5/mlx5_ib.h
index 640840dfc4f4..b2c1005c4373 100644
--- a/drivers/infiniband/hw/mlx5/mlx5_ib.h
+++ b/drivers/infiniband/hw/mlx5/mlx5_ib.h
@@ -392,6 +392,14 @@ struct mlx5_ib_cq {
 	struct ib_umem	       *resize_umem;
 	int			cqe_size;
 	u32			create_flags;
+	struct list_head	wc_list;
+	enum ib_cq_notify_flags notify_flags;
+	struct work_struct	notify_work;
+};
+
+struct mlx5_ib_wc {
+	struct ib_wc wc;
+	struct list_head list;
 };
 
 struct mlx5_ib_srq {
@@ -784,6 +792,8 @@ int mlx5_ib_gsi_post_recv(struct ib_qp *qp, struct ib_recv_wr *wr,
 			  struct ib_recv_wr **bad_wr);
 void mlx5_ib_gsi_pkey_change(struct mlx5_ib_gsi_qp *gsi);
 
+int mlx5_ib_generate_wc(struct ib_cq *ibcq, struct ib_wc *wc);
+
 static inline void init_query_mad(struct ib_smp *mad)
 {
 	mad->base_version  = 1;
-- 
1.7.11.2

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH for-next 07/10] IB/mlx5: Reorder GSI completions

[Haggai Eran]

The emulated GSI QP's send completions are generated by multiple hardware
QPs, so their completions could arrive out of order with respect to the
order their work request were submitted.

Reorder the completions by keeping a list of the posted work request and
their completions. A newly received completion from the hardware updates
the list and marks its work request as completed. However, the completions
are only reported to the client according to the list order.

In order to support that, create a new private CQ to handle the hardware
completions.

Signed-off-by: Haggai Eran <haggaie-VPRAkNaXOzVWk0Htik3J/w@public.gmane.org>
---
 drivers/infiniband/hw/mlx5/gsi.c | 156 +++++++++++++++++++++++++++++++++++++--
 1 file changed, 150 insertions(+), 6 deletions(-)

diff --git a/drivers/infiniband/hw/mlx5/gsi.c b/drivers/infiniband/hw/mlx5/gsi.c
index 1648f539c836..8d040626abb2 100644
--- a/drivers/infiniband/hw/mlx5/gsi.c
+++ b/drivers/infiniband/hw/mlx5/gsi.c
@@ -32,6 +32,13 @@
 
 #include "mlx5_ib.h"
 
+struct mlx5_ib_gsi_wr {
+	struct ib_cqe cqe;
+	struct ib_wc wc;
+	int send_flags;
+	bool completed:1;
+};
+
 struct mlx5_ib_gsi_qp {
 	struct ib_qp ibqp;
 	struct ib_qp *rx_qp;
@@ -40,9 +47,13 @@ struct mlx5_ib_gsi_qp {
 	enum ib_sig_type sq_sig_type;
 	/* Serialize qp state modifications */
 	struct mutex mutex;
+	struct ib_cq *cq;
+	struct mlx5_ib_gsi_wr *outstanding_wrs;
+	u32 outstanding_pi, outstanding_ci;
 	int num_qps;
 	/* Protects access to the tx_qps. Post send operations synchronize
-	 * with tx_qp creation in setup_qp().
+	 * with tx_qp creation in setup_qp(). Also protects the
+	 * outstanding_wrs array and indices.
 	 */
 	spinlock_t lock;
 	struct ib_qp **tx_qps;
@@ -58,6 +69,57 @@ static bool mlx5_ib_deth_sqpn_cap(struct mlx5_ib_dev *dev)
 	return MLX5_CAP_GEN(dev->mdev, set_deth_sqpn);
 }
 
+static u32 next_outstanding(struct mlx5_ib_gsi_qp *gsi, u32 index)
+{
+	return ++index % gsi->cap.max_send_wr;
+}
+
+#define for_each_outstanding_wr(gsi, index) \
+	for (index = gsi->outstanding_ci; index != gsi->outstanding_pi; \
+	     index = next_outstanding(gsi, index))
+
+/* Call with gsi->lock locked */
+static void generate_completions(struct mlx5_ib_gsi_qp *gsi)
+{
+	struct ib_cq *gsi_cq = gsi->ibqp.send_cq;
+	struct mlx5_ib_gsi_wr *wr;
+	u32 index;
+
+	for_each_outstanding_wr(gsi, index) {
+		wr = &gsi->outstanding_wrs[index];
+
+		if (!wr->completed)
+			break;
+
+		if (gsi->sq_sig_type == IB_SIGNAL_ALL_WR ||
+		    wr->send_flags & IB_SEND_SIGNALED)
+			WARN_ON_ONCE(mlx5_ib_generate_wc(gsi_cq, &wr->wc));
+
+		wr->completed = false;
+	}
+
+	gsi->outstanding_ci = index;
+}
+
+static void handle_single_completion(struct ib_cq *cq, struct ib_wc *wc)
+{
+	struct mlx5_ib_gsi_qp *gsi = cq->cq_context;
+	struct mlx5_ib_gsi_wr *wr =
+		container_of(wc->wr_cqe, struct mlx5_ib_gsi_wr, cqe);
+	u64 wr_id;
+	unsigned long flags;
+
+	spin_lock_irqsave(&gsi->lock, flags);
+	wr->completed = true;
+	wr_id = wr->wc.wr_id;
+	wr->wc = *wc;
+	wr->wc.wr_id = wr_id;
+	wr->wc.qp = &gsi->ibqp;
+
+	generate_completions(gsi);
+	spin_unlock_irqrestore(&gsi->lock, flags);
+}
+
 struct ib_qp *mlx5_ib_gsi_create_qp(struct ib_pd *pd,
 				    struct ib_qp_init_attr *init_attr)
 {
@@ -88,6 +150,14 @@ struct ib_qp *mlx5_ib_gsi_create_qp(struct ib_pd *pd,
 		goto err_free;
 	}
 
+	gsi->outstanding_wrs = kcalloc(init_attr->cap.max_send_wr,
+				       sizeof(*gsi->outstanding_wrs),
+				       GFP_KERNEL);
+	if (!gsi->outstanding_wrs) {
+		ret = -ENOMEM;
+		goto err_free_tx;
+	}
+
 	mutex_init(&gsi->mutex);
 
 	mutex_lock(&dev->devr.mutex);
@@ -96,7 +166,7 @@ struct ib_qp *mlx5_ib_gsi_create_qp(struct ib_pd *pd,
 		mlx5_ib_warn(dev, "GSI QP already exists on port %d\n",
 			     port_num);
 		ret = -EBUSY;
-		goto err_free_tx;
+		goto err_free_wrs;
 	}
 	gsi->num_qps = num_qps;
 	spin_lock_init(&gsi->lock);
@@ -106,13 +176,23 @@ struct ib_qp *mlx5_ib_gsi_create_qp(struct ib_pd *pd,
 	gsi->ibqp.qp_num = 1;
 	gsi->port_num = port_num;
 
+	gsi->cq = ib_alloc_cq(pd->device, gsi, init_attr->cap.max_send_wr, 0,
+			      IB_POLL_SOFTIRQ);
+	if (IS_ERR(gsi->cq)) {
+		mlx5_ib_warn(dev, "unable to create send CQ for GSI QP. error %ld\n",
+			     PTR_ERR(gsi->cq));
+		ret = PTR_ERR(gsi->cq);
+		goto err_free_wrs;
+	}
+
 	hw_init_attr.qp_type = MLX5_IB_QPT_HW_GSI;
+	hw_init_attr.send_cq = gsi->cq;
 	gsi->rx_qp = ib_create_qp(pd, &hw_init_attr);
 	if (IS_ERR(gsi->rx_qp)) {
 		mlx5_ib_warn(dev, "unable to create hardware GSI QP. error %ld\n",
 			     PTR_ERR(gsi->rx_qp));
 		ret = PTR_ERR(gsi->rx_qp);
-		goto err_free_tx;
+		goto err_destroy_cq;
 	}
 
 	dev->devr.ports[init_attr->port_num - 1].gsi = gsi;
@@ -121,8 +201,12 @@ struct ib_qp *mlx5_ib_gsi_create_qp(struct ib_pd *pd,
 
 	return &gsi->ibqp;
 
-err_free_tx:
+err_destroy_cq:
+	ib_free_cq(gsi->cq);
+err_free_wrs:
 	mutex_unlock(&dev->devr.mutex);
+	kfree(gsi->outstanding_wrs);
+err_free_tx:
 	kfree(gsi->tx_qps);
 err_free:
 	kfree(gsi);
@@ -158,6 +242,9 @@ int mlx5_ib_gsi_destroy_qp(struct ib_qp *qp)
 		gsi->tx_qps[qp_index] = NULL;
 	}
 
+	ib_free_cq(gsi->cq);
+
+	kfree(gsi->outstanding_wrs);
 	kfree(gsi->tx_qps);
 	kfree(gsi);
 
@@ -170,7 +257,7 @@ static struct ib_qp *create_gsi_ud_qp(struct mlx5_ib_gsi_qp *gsi)
 	struct ib_qp_init_attr init_attr = {
 		.event_handler = gsi->rx_qp->event_handler,
 		.qp_context = gsi->rx_qp->qp_context,
-		.send_cq = gsi->rx_qp->send_cq,
+		.send_cq = gsi->cq,
 		.recv_cq = gsi->rx_qp->recv_cq,
 		.cap = {
 			.max_send_wr = gsi->cap.max_send_wr,
@@ -326,12 +413,69 @@ int mlx5_ib_gsi_query_qp(struct ib_qp *qp, struct ib_qp_attr *qp_attr,
 	return ret;
 }
 
+/* Call with gsi->lock locked */
+static int mlx5_ib_add_outstanding_wr(struct mlx5_ib_gsi_qp *gsi,
+				      struct ib_ud_wr *wr, struct ib_wc *wc)
+{
+	struct mlx5_ib_dev *dev = to_mdev(gsi->rx_qp->device);
+	struct mlx5_ib_gsi_wr *gsi_wr;
+
+	if (gsi->outstanding_pi == gsi->outstanding_ci + gsi->cap.max_send_wr) {
+		mlx5_ib_warn(dev, "no available GSI work request.\n");
+		return -ENOMEM;
+	}
+
+	gsi_wr = &gsi->outstanding_wrs[gsi->outstanding_pi];
+	gsi->outstanding_pi = next_outstanding(gsi, gsi->outstanding_pi);
+
+	if (!wc) {
+		memset(&gsi_wr->wc, 0, sizeof(gsi_wr->wc));
+		gsi_wr->wc.pkey_index = wr->pkey_index;
+		gsi_wr->wc.wr_id = wr->wr.wr_id;
+	} else {
+		gsi_wr->wc = *wc;
+		gsi_wr->completed = true;
+	}
+
+	gsi_wr->cqe.done = &handle_single_completion;
+	wr->wr.wr_cqe = &gsi_wr->cqe;
+
+	return 0;
+}
+
 int mlx5_ib_gsi_post_send(struct ib_qp *qp, struct ib_send_wr *wr,
 			  struct ib_send_wr **bad_wr)
 {
 	struct mlx5_ib_gsi_qp *gsi = gsi_qp(qp);
+	unsigned long flags;
+	int ret;
+
+	for (; wr; wr = wr->next) {
+		struct ib_ud_wr cur_wr = *ud_wr(wr);
+
+		cur_wr.wr.next = NULL;
 
-	return ib_post_send(gsi->rx_qp, wr, bad_wr);
+		spin_lock_irqsave(&gsi->lock, flags);
+		ret = mlx5_ib_add_outstanding_wr(gsi, &cur_wr, NULL);
+		if (ret)
+			goto err;
+
+		ret = ib_post_send(gsi->rx_qp, &cur_wr.wr, bad_wr);
+		if (ret) {
+			/* Undo the effect of adding the outstanding wr */
+			gsi->outstanding_pi = (gsi->outstanding_pi - 1) %
+					      gsi->cap.max_send_wr;
+			goto err;
+		}
+		spin_unlock_irqrestore(&gsi->lock, flags);
+	}
+
+	return 0;
+
+err:
+	spin_unlock_irqrestore(&gsi->lock, flags);
+	*bad_wr = wr;
+	return ret;
 }
 
 int mlx5_ib_gsi_post_recv(struct ib_qp *qp, struct ib_recv_wr *wr,
-- 
1.7.11.2

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH for-next 08/10] IB/mlx5: Pick the right GSI transmission QP for sending

[Haggai Eran]

Pick the QP to use according to the wr.ud.pkey_index field in the work
request. If the QP doesn't exist, it means the P_Key is zero and the packet
would have been dropped, so just generate a completion and move on.

Reviewed-by: Leon Romanovsky <leonro-VPRAkNaXOzVWk0Htik3J/w@public.gmane.org>
Signed-off-by: Haggai Eran <haggaie-VPRAkNaXOzVWk0Htik3J/w@public.gmane.org>
---
 drivers/infiniband/hw/mlx5/gsi.c | 48 +++++++++++++++++++++++++++++++++++++++-
 1 file changed, 47 insertions(+), 1 deletion(-)

diff --git a/drivers/infiniband/hw/mlx5/gsi.c b/drivers/infiniband/hw/mlx5/gsi.c
index 8d040626abb2..938f6ddca4b3 100644
--- a/drivers/infiniband/hw/mlx5/gsi.c
+++ b/drivers/infiniband/hw/mlx5/gsi.c
@@ -443,10 +443,47 @@ static int mlx5_ib_add_outstanding_wr(struct mlx5_ib_gsi_qp *gsi,
 	return 0;
 }
 
+/* Call with gsi->lock locked */
+static int mlx5_ib_gsi_silent_drop(struct mlx5_ib_gsi_qp *gsi,
+				    struct ib_ud_wr *wr)
+{
+	struct ib_wc wc = {
+		{ .wr_id = wr->wr.wr_id },
+		.status = IB_WC_SUCCESS,
+		.opcode = IB_WC_SEND,
+		.qp = &gsi->ibqp,
+	};
+	int ret;
+
+	ret = mlx5_ib_add_outstanding_wr(gsi, wr, &wc);
+	if (ret)
+		return ret;
+
+	generate_completions(gsi);
+
+	return 0;
+}
+
+/* Call with gsi->lock locked */
+static struct ib_qp *get_tx_qp(struct mlx5_ib_gsi_qp *gsi, struct ib_ud_wr *wr)
+{
+	struct mlx5_ib_dev *dev = to_mdev(gsi->rx_qp->device);
+	int qp_index = wr->pkey_index;
+
+	if (!mlx5_ib_deth_sqpn_cap(dev))
+		return gsi->rx_qp;
+
+	if (qp_index >= gsi->num_qps)
+		return NULL;
+
+	return gsi->tx_qps[qp_index];
+}
+
 int mlx5_ib_gsi_post_send(struct ib_qp *qp, struct ib_send_wr *wr,
 			  struct ib_send_wr **bad_wr)
 {
 	struct mlx5_ib_gsi_qp *gsi = gsi_qp(qp);
+	struct ib_qp *tx_qp;
 	unsigned long flags;
 	int ret;
 
@@ -456,11 +493,20 @@ int mlx5_ib_gsi_post_send(struct ib_qp *qp, struct ib_send_wr *wr,
 		cur_wr.wr.next = NULL;
 
 		spin_lock_irqsave(&gsi->lock, flags);
+		tx_qp = get_tx_qp(gsi, &cur_wr);
+		if (!tx_qp) {
+			ret = mlx5_ib_gsi_silent_drop(gsi, &cur_wr);
+			if (ret)
+				goto err;
+			spin_unlock_irqrestore(&gsi->lock, flags);
+			continue;
+		}
+
 		ret = mlx5_ib_add_outstanding_wr(gsi, &cur_wr, NULL);
 		if (ret)
 			goto err;
 
-		ret = ib_post_send(gsi->rx_qp, &cur_wr.wr, bad_wr);
+		ret = ib_post_send(tx_qp, &cur_wr.wr, bad_wr);
 		if (ret) {
 			/* Undo the effect of adding the outstanding wr */
 			gsi->outstanding_pi = (gsi->outstanding_pi - 1) %
-- 
1.7.11.2

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH for-next 09/10] IB/mlx5: Eliminate GSI RX QP's send buffers

[Haggai Eran]

Now that the transmission of GSI MADs is done with the special transmission
QPs, eliminate the send buffers in the GSI receive QP.

Reviewed-by: Leon Romanovsky <leonro-VPRAkNaXOzVWk0Htik3J/w@public.gmane.org>
Signed-off-by: Haggai Eran <haggaie-VPRAkNaXOzVWk0Htik3J/w@public.gmane.org>
---
 drivers/infiniband/hw/mlx5/gsi.c | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/drivers/infiniband/hw/mlx5/gsi.c b/drivers/infiniband/hw/mlx5/gsi.c
index 938f6ddca4b3..53e03c8ede79 100644
--- a/drivers/infiniband/hw/mlx5/gsi.c
+++ b/drivers/infiniband/hw/mlx5/gsi.c
@@ -187,6 +187,11 @@ struct ib_qp *mlx5_ib_gsi_create_qp(struct ib_pd *pd,
 
 	hw_init_attr.qp_type = MLX5_IB_QPT_HW_GSI;
 	hw_init_attr.send_cq = gsi->cq;
+	if (num_qps) {
+		hw_init_attr.cap.max_send_wr = 0;
+		hw_init_attr.cap.max_send_sge = 0;
+		hw_init_attr.cap.max_inline_data = 0;
+	}
 	gsi->rx_qp = ib_create_qp(pd, &hw_init_attr);
 	if (IS_ERR(gsi->rx_qp)) {
 		mlx5_ib_warn(dev, "unable to create hardware GSI QP. error %ld\n",
-- 
1.7.11.2

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH for-next 10/10] IB/cma: Print warning on different inner and header P_Keys

[Haggai Eran]

Commit 4c21b5bcef73 ("IB/cma: Add net_dev and private data checks to RDMA
CM") added checks for incoming RDMA CM requests that they can be matched to
a netdev based on the P_Key in the BTH of the request. This behavior was
reverted in commit ab3964ad2acf ("IB/cma: Use inner P_Key to determine
netdev"), since the mlx5 and ipath drivers didn't send the correct value
in the BTH P_Key.

Since the ipath driver was removed, and the mlx5 driver can now send GSI
packets on different P_Keys, we could revert the patch to let the rdma_cm
module look on the BTH P_Key when deciding to what netdev a packet belongs.
However, that still breaks compatibility with the older drivers.

Change the behavior to print a warning when receiving a request that has a
different BTH P_Key and inner payload P_Key. In the future, after users
have seen the warnings and upgraded their setups, remove the warning and
block these requests.

Signed-off-by: Haggai Eran <haggaie-VPRAkNaXOzVWk0Htik3J/w@public.gmane.org>
---
 drivers/infiniband/core/cma.c | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/drivers/infiniband/core/cma.c b/drivers/infiniband/core/cma.c
index 9729639df407..7eace1ffa74d 100644
--- a/drivers/infiniband/core/cma.c
+++ b/drivers/infiniband/core/cma.c
@@ -1206,6 +1206,10 @@ static int cma_save_req_info(const struct ib_cm_event *ib_event,
 		req->has_gid	= true;
 		req->service_id	= req_param->primary_path->service_id;
 		req->pkey	= be16_to_cpu(req_param->primary_path->pkey);
+		if (req->pkey != req_param->bth_pkey)
+			pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and primary path P_Key (0x%x)\n"
+					    "RDMA CMA: in the future this may cause the request to be dropped\n",
+					    req_param->bth_pkey, req->pkey);
 		break;
 	case IB_CM_SIDR_REQ_RECEIVED:
 		req->device	= sidr_param->listen_id->device;
@@ -1213,6 +1217,10 @@ static int cma_save_req_info(const struct ib_cm_event *ib_event,
 		req->has_gid	= false;
 		req->service_id	= sidr_param->service_id;
 		req->pkey	= sidr_param->pkey;
+		if (req->pkey != sidr_param->bth_pkey)
+			pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and SIDR request payload P_Key (0x%x)\n"
+					    "RDMA CMA: in the future this may cause the request to be dropped\n",
+					    sidr_param->bth_pkey, req->pkey);
 		break;
 	default:
 		return -EINVAL;
-- 
1.7.11.2

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [PATCH for-next 00/10] IB/mlx5: Use wr.ud.pkey_index in GSI QP

[Doug Ledford]

[-- Attachment #1: Type: text/plain, Size: 3594 bytes --]

On 02/29/2016 08:45 AM, Haggai Eran wrote:
> The check of BTH P_Key for RDMA_CM requests was introduced in patch
> 4c21b5bcef73 ("IB/cma: Add net_dev and private data checks to RDMA CM").
> A side effect of that change was the unexpected behavior of the mlx5 driver
> in handling the pkey_index in ib_send_wr.ud in GSI packets.  The pkey_index
> in ib_wc was not set correctly [1], and in addition, we found out that the
> mlx5 driver did not respect the pkey_index in ib_send_wr.ud for GSI
> packets. Since the pkey_index in a Work Request is not explicitly required
> by the IBA specifications, the Connect-IB and ConnectX-4 adapter cards do
> not support it.
> 
> This patch fixes the issue inside the mlx5_ib module. It emulates a single
> GSI QP by creating multiple QPs underneath the surface. An ordinary GSI QP
> is created for receiving MADs destined at QP 1, and multiple transmission
> QPs are created for sending MADs. These QPs behave like UD QPs, although
> they use QP 1 as their source QP number. The QPs are created on-demand, one
> for each non-zero P_Key in the P_Key table.
> 
> The driver handles reordering of the completions to make sure they are
> reported in the same order their matching Work Requests were posted.
> 
> A spinlock in the emulated GSI QP struct protects concurrent access to the
> transmission QPs from post send operations and from their creation
> following a P_Key table change. In addition, it protects the Work Request
> list that maintains the completion order.
> 
> The patchset is divided as follows. The first patch adds the infrastructure
> required for creating additional QPs that send as QP 1. The following patch
> adds some helpful debugging prints to the QP modification code. Patch 3
> adds a skeleton code for an emulated GSI QP. The next two patches create
> the multiple transmission QPs. Patches 6-7 handle the reordering of the
> completions. Patches 8-9 enable the use of the created QPs and remove the
> send queue buffer from the receive GSI QP. Finally patch 10 reverts the
> change in "IB/cma: Use inner P_Key to determine netdev".
> 
> These patches apply against kernel v4.5-rc6.
> 
> [1] mlx5: Fix incorrect wc pkey_index assignment for GSI messages
> http://www.spinics.net/lists/linux-rdma/msg28374.html
> 
> Haggai Eran (10):
>   IB/mlx5: Add support for setting source QP number
>   IB/mlx5: Modify QP debugging prints
>   IB/mlx5: Add GSI QP wrapper
>   IB/mlx5: Create multiple transmission GSI QPs
>   IB/mlx5: Create GSI transmission QPs when P_Key table is changed
>   IB/mlx5: Generate completions in software
>   IB/mlx5: Reorder GSI completions
>   IB/mlx5: Pick the right GSI transmission QP for sending
>   IB/mlx5: Eliminate GSI RX QP's send buffers
>   IB/cma: Print warning on different inner and header P_Keys
> 
>  drivers/infiniband/core/cma.c        |   8 +
>  drivers/infiniband/hw/mlx5/Makefile  |   2 +-
>  drivers/infiniband/hw/mlx5/cq.c      |  83 +++++-
>  drivers/infiniband/hw/mlx5/gsi.c     | 548 +++++++++++++++++++++++++++++++++++
>  drivers/infiniband/hw/mlx5/main.c    |  30 ++
>  drivers/infiniband/hw/mlx5/mlx5_ib.h |  57 +++-
>  drivers/infiniband/hw/mlx5/qp.c      |  95 ++++--
>  include/linux/mlx5/mlx5_ifc.h        |   3 +-
>  include/linux/mlx5/qp.h              |   3 +-
>  9 files changed, 802 insertions(+), 27 deletions(-)
>  create mode 100644 drivers/infiniband/hw/mlx5/gsi.c
> 

Thanks, applied.

-- 
Doug Ledford <dledford-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
              GPG KeyID: 0E572FDD



[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 884 bytes --]