From: Maor Gottlieb <maorg@mellanox.com>
To: davem@davemloft.net, jgg@mellanox.com, dledford@redhat.com,
j.vosburgh@gmail.com, vfalico@gmail.com, andy@greyhouse.net,
kuba@kernel.org, jiri@mellanox.com, dsahern@kernel.org
Cc: leonro@mellanox.com, saeedm@mellanox.com,
linux-rdma@vger.kernel.org, netdev@vger.kernel.org,
alexr@mellanox.com, Maor Gottlieb <maorg@mellanox.com>
Subject: [PATCH V6 mlx5-next 16/16] RDMA/mlx5: Set lag tx affinity according to slave
Date: Sun, 26 Apr 2020 10:17:17 +0300 [thread overview]
Message-ID: <20200426071717.17088-17-maorg@mellanox.com> (raw)
In-Reply-To: <20200426071717.17088-1-maorg@mellanox.com>
The patch sets the lag tx affinity of the data QPs and the
GSI QPs according to the LAG xmit slave.
For GSI QPs, in case that the link layer is Ethenet (RoCE) we create
two GSI QPs, one for each physical port. When the driver selects the
GSI QP, it will consider the port affinity result.
For connected QPs, the driver sets the affinity of the xmit slave.
The above, ensure that RC QP and it's corresponding GSI QP will
transmit from the same physical port.
Signed-off-by: Maor Gottlieb <maorg@mellanox.com>
Reviewed-by: Leon Romanovsky <leonro@mellanox.com>
---
drivers/infiniband/hw/mlx5/ah.c | 9 +++--
drivers/infiniband/hw/mlx5/gsi.c | 34 ++++++++++++++----
drivers/infiniband/hw/mlx5/main.c | 2 ++
drivers/infiniband/hw/mlx5/mlx5_ib.h | 1 +
drivers/infiniband/hw/mlx5/qp.c | 52 +++++++++++++++++++---------
include/linux/mlx5/mlx5_ifc.h | 4 ++-
include/linux/mlx5/qp.h | 2 ++
7 files changed, 77 insertions(+), 27 deletions(-)
diff --git a/drivers/infiniband/hw/mlx5/ah.c b/drivers/infiniband/hw/mlx5/ah.c
index 9b59348d51b5..cc858f658567 100644
--- a/drivers/infiniband/hw/mlx5/ah.c
+++ b/drivers/infiniband/hw/mlx5/ah.c
@@ -33,8 +33,9 @@
#include "mlx5_ib.h"
static void create_ib_ah(struct mlx5_ib_dev *dev, struct mlx5_ib_ah *ah,
- struct rdma_ah_attr *ah_attr)
+ struct rdma_ah_init_attr *init_attr)
{
+ struct rdma_ah_attr *ah_attr = init_attr->ah_attr;
enum ib_gid_type gid_type;
if (rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH) {
@@ -51,6 +52,10 @@ static void create_ib_ah(struct mlx5_ib_dev *dev, struct mlx5_ib_ah *ah,
ah->av.stat_rate_sl = (rdma_ah_get_static_rate(ah_attr) << 4);
if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) {
+ if (init_attr->xmit_slave)
+ ah->xmit_port =
+ mlx5_lag_get_slave_port(dev->mdev,
+ init_attr->xmit_slave);
gid_type = ah_attr->grh.sgid_attr->gid_type;
memcpy(ah->av.rmac, ah_attr->roce.dmac,
@@ -98,7 +103,7 @@ int mlx5_ib_create_ah(struct ib_ah *ibah, struct rdma_ah_init_attr *init_attr,
return err;
}
- create_ib_ah(dev, ah, ah_attr);
+ create_ib_ah(dev, ah, init_attr);
return 0;
}
diff --git a/drivers/infiniband/hw/mlx5/gsi.c b/drivers/infiniband/hw/mlx5/gsi.c
index 1ae6fd95acaa..fbae1c094fe2 100644
--- a/drivers/infiniband/hw/mlx5/gsi.c
+++ b/drivers/infiniband/hw/mlx5/gsi.c
@@ -119,12 +119,20 @@ 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 num_qps = 0;
int ret;
mlx5_ib_dbg(dev, "creating GSI QP\n");
+ if (mlx5_ib_deth_sqpn_cap(dev)) {
+ if (MLX5_CAP_GEN(dev->mdev,
+ port_type) == MLX5_CAP_PORT_TYPE_IB)
+ num_qps = pd->device->attrs.max_pkeys;
+ else if (dev->lag_active)
+ num_qps = MLX5_MAX_PORTS;
+ }
+
+
if (port_num > ARRAY_SIZE(dev->devr.ports) || port_num < 1) {
mlx5_ib_warn(dev,
"invalid port number %d during GSI QP creation\n",
@@ -270,7 +278,7 @@ static struct ib_qp *create_gsi_ud_qp(struct mlx5_ib_gsi_qp *gsi)
}
static int modify_to_rts(struct mlx5_ib_gsi_qp *gsi, struct ib_qp *qp,
- u16 qp_index)
+ u16 pkey_index)
{
struct mlx5_ib_dev *dev = to_mdev(qp->device);
struct ib_qp_attr attr;
@@ -279,7 +287,7 @@ static int modify_to_rts(struct mlx5_ib_gsi_qp *gsi, struct ib_qp *qp,
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.pkey_index = pkey_index;
attr.qkey = IB_QP1_QKEY;
attr.port_num = gsi->port_num;
ret = ib_modify_qp(qp, &attr, mask);
@@ -313,12 +321,17 @@ 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);
+ int pkey_index = qp_index;
+ struct mlx5_ib_qp *mqp;
struct ib_qp *qp;
unsigned long flags;
u16 pkey;
int ret;
- ret = ib_query_pkey(device, gsi->port_num, qp_index, &pkey);
+ if (MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_IB)
+ pkey_index = 0;
+
+ ret = ib_query_pkey(device, gsi->port_num, pkey_index, &pkey);
if (ret) {
mlx5_ib_warn(dev, "unable to read P_Key at port %d, index %d\n",
gsi->port_num, qp_index);
@@ -347,7 +360,10 @@ static void setup_qp(struct mlx5_ib_gsi_qp *gsi, u16 qp_index)
return;
}
- ret = modify_to_rts(gsi, qp, qp_index);
+ mqp = to_mqp(qp);
+ if (dev->lag_active)
+ mqp->gsi_lag_port = qp_index + 1;
+ ret = modify_to_rts(gsi, qp, pkey_index);
if (ret)
goto err_destroy_qp;
@@ -466,11 +482,15 @@ static int mlx5_ib_gsi_silent_drop(struct mlx5_ib_gsi_qp *gsi,
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);
+ struct mlx5_ib_ah *ah = to_mah(wr->ah);
int qp_index = wr->pkey_index;
- if (!mlx5_ib_deth_sqpn_cap(dev))
+ if (!gsi->num_qps)
return gsi->rx_qp;
+ if (dev->lag_active && ah->xmit_port)
+ qp_index = ah->xmit_port - 1;
+
if (qp_index >= gsi->num_qps)
return NULL;
diff --git a/drivers/infiniband/hw/mlx5/main.c b/drivers/infiniband/hw/mlx5/main.c
index 6679756506e6..2db2309dde47 100644
--- a/drivers/infiniband/hw/mlx5/main.c
+++ b/drivers/infiniband/hw/mlx5/main.c
@@ -53,6 +53,7 @@
#include <linux/list.h>
#include <rdma/ib_smi.h>
#include <rdma/ib_umem.h>
+#include <rdma/lag.h>
#include <linux/in.h>
#include <linux/etherdevice.h>
#include "mlx5_ib.h"
@@ -6549,6 +6550,7 @@ static int mlx5_ib_stage_init_init(struct mlx5_ib_dev *dev)
dev->ib_dev.phys_port_cnt = dev->num_ports;
dev->ib_dev.num_comp_vectors = mlx5_comp_vectors_count(mdev);
dev->ib_dev.dev.parent = mdev->device;
+ dev->ib_dev.lag_flags = RDMA_LAG_FLAGS_HASH_ALL_SLAVES;
mutex_init(&dev->cap_mask_mutex);
INIT_LIST_HEAD(&dev->qp_list);
diff --git a/drivers/infiniband/hw/mlx5/mlx5_ib.h b/drivers/infiniband/hw/mlx5/mlx5_ib.h
index 524c188cb4b3..a924ac96b32f 100644
--- a/drivers/infiniband/hw/mlx5/mlx5_ib.h
+++ b/drivers/infiniband/hw/mlx5/mlx5_ib.h
@@ -471,6 +471,7 @@ struct mlx5_ib_qp {
* but not take effective
*/
u32 counter_pending;
+ u16 gsi_lag_port;
};
struct mlx5_ib_cq_buf {
diff --git a/drivers/infiniband/hw/mlx5/qp.c b/drivers/infiniband/hw/mlx5/qp.c
index a45499809903..7fd1d4f81862 100644
--- a/drivers/infiniband/hw/mlx5/qp.c
+++ b/drivers/infiniband/hw/mlx5/qp.c
@@ -3052,10 +3052,12 @@ static enum mlx5_qp_optpar opt_mask[MLX5_QP_NUM_STATE][MLX5_QP_NUM_STATE][MLX5_Q
MLX5_QP_OPTPAR_RAE |
MLX5_QP_OPTPAR_RWE |
MLX5_QP_OPTPAR_PKEY_INDEX |
- MLX5_QP_OPTPAR_PRI_PORT,
+ MLX5_QP_OPTPAR_PRI_PORT |
+ MLX5_QP_OPTPAR_LAG_TX_AFF,
[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE |
MLX5_QP_OPTPAR_PKEY_INDEX |
- MLX5_QP_OPTPAR_PRI_PORT,
+ MLX5_QP_OPTPAR_PRI_PORT |
+ MLX5_QP_OPTPAR_LAG_TX_AFF,
[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX |
MLX5_QP_OPTPAR_Q_KEY |
MLX5_QP_OPTPAR_PRI_PORT,
@@ -3063,17 +3065,20 @@ static enum mlx5_qp_optpar opt_mask[MLX5_QP_NUM_STATE][MLX5_QP_NUM_STATE][MLX5_Q
MLX5_QP_OPTPAR_RAE |
MLX5_QP_OPTPAR_RWE |
MLX5_QP_OPTPAR_PKEY_INDEX |
- MLX5_QP_OPTPAR_PRI_PORT,
+ MLX5_QP_OPTPAR_PRI_PORT |
+ MLX5_QP_OPTPAR_LAG_TX_AFF,
},
[MLX5_QP_STATE_RTR] = {
[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH |
MLX5_QP_OPTPAR_RRE |
MLX5_QP_OPTPAR_RAE |
MLX5_QP_OPTPAR_RWE |
- MLX5_QP_OPTPAR_PKEY_INDEX,
+ MLX5_QP_OPTPAR_PKEY_INDEX |
+ MLX5_QP_OPTPAR_LAG_TX_AFF,
[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH |
MLX5_QP_OPTPAR_RWE |
- MLX5_QP_OPTPAR_PKEY_INDEX,
+ MLX5_QP_OPTPAR_PKEY_INDEX |
+ MLX5_QP_OPTPAR_LAG_TX_AFF,
[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX |
MLX5_QP_OPTPAR_Q_KEY,
[MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_PKEY_INDEX |
@@ -3082,7 +3087,8 @@ static enum mlx5_qp_optpar opt_mask[MLX5_QP_NUM_STATE][MLX5_QP_NUM_STATE][MLX5_Q
MLX5_QP_OPTPAR_RRE |
MLX5_QP_OPTPAR_RAE |
MLX5_QP_OPTPAR_RWE |
- MLX5_QP_OPTPAR_PKEY_INDEX,
+ MLX5_QP_OPTPAR_PKEY_INDEX |
+ MLX5_QP_OPTPAR_LAG_TX_AFF,
},
},
[MLX5_QP_STATE_RTR] = {
@@ -3435,11 +3441,8 @@ static unsigned int get_tx_affinity_rr(struct mlx5_ib_dev *dev,
static bool qp_supports_affinity(struct ib_qp *qp)
{
- struct mlx5_ib_qp *mqp = to_mqp(qp);
-
if ((qp->qp_type == IB_QPT_RC) ||
- (qp->qp_type == IB_QPT_UD &&
- !(mqp->flags & MLX5_IB_QP_SQPN_QP1)) ||
+ (qp->qp_type == IB_QPT_UD) ||
(qp->qp_type == IB_QPT_UC) ||
(qp->qp_type == IB_QPT_RAW_PACKET) ||
(qp->qp_type == IB_QPT_XRC_INI) ||
@@ -3448,7 +3451,9 @@ static bool qp_supports_affinity(struct ib_qp *qp)
return false;
}
-static unsigned int get_tx_affinity(struct ib_qp *qp, u8 init,
+static unsigned int get_tx_affinity(struct ib_qp *qp,
+ const struct ib_qp_attr *attr,
+ int attr_mask, u8 init,
struct ib_udata *udata)
{
struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
@@ -3458,10 +3463,18 @@ static unsigned int get_tx_affinity(struct ib_qp *qp, u8 init,
struct mlx5_ib_qp_base *qp_base;
unsigned int tx_affinity;
- if (!(dev->lag_active && init && qp_supports_affinity(qp)))
+ if (!(dev->lag_active && qp_supports_affinity(qp)))
return 0;
- tx_affinity = get_tx_affinity_rr(dev, udata);
+ if (mqp->flags & MLX5_IB_QP_SQPN_QP1)
+ tx_affinity = mqp->gsi_lag_port;
+ else if (init)
+ tx_affinity = get_tx_affinity_rr(dev, udata);
+ else if ((attr_mask & IB_QP_AV) && attr->xmit_slave)
+ tx_affinity =
+ mlx5_lag_get_slave_port(dev->mdev, attr->xmit_slave);
+ else
+ return 0;
qp_base = &mqp->trans_qp.base;
if (ucontext)
@@ -3547,7 +3560,7 @@ static int __mlx5_ib_modify_qp(struct ib_qp *ibqp,
struct mlx5_qp_context *context;
struct mlx5_ib_pd *pd;
enum mlx5_qp_state mlx5_cur, mlx5_new;
- enum mlx5_qp_optpar optpar;
+ enum mlx5_qp_optpar optpar = 0;
u32 set_id = 0;
int mlx5_st;
int err;
@@ -3582,10 +3595,15 @@ static int __mlx5_ib_modify_qp(struct ib_qp *ibqp,
}
}
- tx_affinity = get_tx_affinity(ibqp,
+ tx_affinity = get_tx_affinity(ibqp, attr, attr_mask,
cur_state == IB_QPS_RESET &&
new_state == IB_QPS_INIT, udata);
- context->flags |= cpu_to_be32(tx_affinity << 24);
+ if (tx_affinity) {
+ context->flags |= cpu_to_be32(tx_affinity << 24);
+ if (new_state == IB_QPS_RTR &&
+ MLX5_CAP_GEN(dev->mdev, init2_lag_tx_port_affinity))
+ optpar |= MLX5_QP_OPTPAR_LAG_TX_AFF;
+ }
if (is_sqp(ibqp->qp_type)) {
context->mtu_msgmax = (IB_MTU_256 << 5) | 8;
@@ -3722,7 +3740,7 @@ static int __mlx5_ib_modify_qp(struct ib_qp *ibqp,
}
op = optab[mlx5_cur][mlx5_new];
- optpar = ib_mask_to_mlx5_opt(attr_mask);
+ optpar |= ib_mask_to_mlx5_opt(attr_mask);
optpar &= opt_mask[mlx5_cur][mlx5_new][mlx5_st];
if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET ||
diff --git a/include/linux/mlx5/mlx5_ifc.h b/include/linux/mlx5/mlx5_ifc.h
index 69b27c7dfc3e..a3b6c92e889e 100644
--- a/include/linux/mlx5/mlx5_ifc.h
+++ b/include/linux/mlx5/mlx5_ifc.h
@@ -1296,7 +1296,9 @@ struct mlx5_ifc_cmd_hca_cap_bits {
u8 wol_p[0x1];
u8 stat_rate_support[0x10];
- u8 reserved_at_1f0[0xc];
+ u8 reserved_at_1f0[0x8];
+ u8 init2_lag_tx_port_affinity[0x1];
+ u8 reserved_at_1f9[0x3];
u8 cqe_version[0x4];
u8 compact_address_vector[0x1];
diff --git a/include/linux/mlx5/qp.h b/include/linux/mlx5/qp.h
index ae63b1ae9004..fab88b0c76f9 100644
--- a/include/linux/mlx5/qp.h
+++ b/include/linux/mlx5/qp.h
@@ -66,6 +66,7 @@ enum mlx5_qp_optpar {
MLX5_QP_OPTPAR_RETRY_COUNT = 1 << 12,
MLX5_QP_OPTPAR_RNR_RETRY = 1 << 13,
MLX5_QP_OPTPAR_ACK_TIMEOUT = 1 << 14,
+ MLX5_QP_OPTPAR_LAG_TX_AFF = 1 << 15,
MLX5_QP_OPTPAR_PRI_PORT = 1 << 16,
MLX5_QP_OPTPAR_SRQN = 1 << 18,
MLX5_QP_OPTPAR_CQN_RCV = 1 << 19,
@@ -315,6 +316,7 @@ struct mlx5_av {
struct mlx5_ib_ah {
struct ib_ah ibah;
struct mlx5_av av;
+ u8 xmit_port;
};
static inline struct mlx5_ib_ah *to_mah(struct ib_ah *ibah)
--
2.17.2
next prev parent reply other threads:[~2020-04-26 7:17 UTC|newest]
Thread overview: 31+ messages / expand[flat|nested] mbox.gz Atom feed top
2020-04-26 7:17 [PATCH 00/16] Add support to get xmit slave Maor Gottlieb
2020-04-26 7:17 ` [PATCH V6 mlx5-next 01/16] net/core: Introduce netdev_get_xmit_slave Maor Gottlieb
2020-04-26 16:03 ` David Ahern
2020-04-26 7:17 ` [PATCH V6 mlx5-next 02/16] bonding: Export skip slave logic to function Maor Gottlieb
2020-04-26 7:17 ` [PATCH V6 mlx5-next 03/16] bonding: Rename slave_arr to usable_slaves Maor Gottlieb
2020-04-26 7:17 ` [PATCH V6 mlx5-next 04/16] bonding/alb: Add helper functions to get the xmit slave Maor Gottlieb
2020-04-26 7:17 ` [PATCH 05/16] bonding: Add helper function to get the xmit slave based on hash Maor Gottlieb
2020-04-26 7:17 ` [PATCH V6 mlx5-next 06/16] bonding: Add helper function to get the xmit slave in rr mode Maor Gottlieb
2020-04-26 7:17 ` [PATCH V6 mlx5-next 07/16] bonding: Add function to get the xmit slave in active-backup mode Maor Gottlieb
2020-04-26 7:17 ` [PATCH V6 mlx5-next 08/16] bonding: Add array of all slaves Maor Gottlieb
2020-04-26 7:17 ` [PATCH V6 mlx5-next 09/16] bonding: Implement ndo_get_xmit_slave Maor Gottlieb
2020-04-26 7:17 ` [PATCH V6 mlx5-next 10/16] RDMA: Group create AH arguments in struct Maor Gottlieb
2020-04-28 8:39 ` Devesh Sharma
2020-04-28 10:17 ` Gal Pressman
2020-04-28 10:51 ` liweihang
2020-04-28 23:15 ` Jason Gunthorpe
2020-04-26 7:17 ` [PATCH V6 mlx5-next 11/16] RDMA/core: Add LAG functionality Maor Gottlieb
2020-04-28 23:15 ` Jason Gunthorpe
2020-04-28 23:30 ` Jason Gunthorpe
2020-04-29 9:01 ` Maor Gottlieb
2020-04-29 15:01 ` Jason Gunthorpe
2020-04-26 7:17 ` [PATCH V6 mlx5-next 12/16] RDMA/core: Get xmit slave for LAG Maor Gottlieb
2020-04-28 23:33 ` Jason Gunthorpe
2020-04-26 7:17 ` [PATCH V6 mlx5-next 13/16] net/mlx5: Change lag mutex lock to spin lock Maor Gottlieb
2020-04-28 23:23 ` Jason Gunthorpe
2020-04-26 7:17 ` [PATCH V6 mlx5-next 14/16] net/mlx5: Add support to get lag physical port Maor Gottlieb
2020-04-26 7:17 ` [PATCH V6 mlx5-next 15/16] RDMA/mlx5: Refactor affinity related code Maor Gottlieb
2020-04-26 7:17 ` Maor Gottlieb [this message]
2020-04-26 7:37 ` [PATCH 00/16] Add support to get xmit slave Maor Gottlieb
2020-04-28 20:04 ` Saeed Mahameed
2020-04-28 21:37 ` David Miller
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20200426071717.17088-17-maorg@mellanox.com \
--to=maorg@mellanox.com \
--cc=alexr@mellanox.com \
--cc=andy@greyhouse.net \
--cc=davem@davemloft.net \
--cc=dledford@redhat.com \
--cc=dsahern@kernel.org \
--cc=j.vosburgh@gmail.com \
--cc=jgg@mellanox.com \
--cc=jiri@mellanox.com \
--cc=kuba@kernel.org \
--cc=leonro@mellanox.com \
--cc=linux-rdma@vger.kernel.org \
--cc=netdev@vger.kernel.org \
--cc=saeedm@mellanox.com \
--cc=vfalico@gmail.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.