Add ConnectX DCS offload support
Subsystems (5)
cover Add ConnectX DCS offload support
From: Leon Romanovsky <[email protected]>
Changelog:
v1:
* Rephrase commit message of second patch
v0: https://lore.kernel.org/linux-rdma/[email protected]
------------
This patchset from Lior adds support of DCI stream channel (DCS) support.
DCS is an offload to SW load balancing of DC initiator work requests.
A single DC QP initiator (DCI) can be connected to only one target at the time and can't start new connection until the previous work request is completed.
This limitation causes to delays when the initiator process needs to transfer data to multiple targets at the same time.
Thanks
Lior Nahmanson (3):
net/mlx5: Add DCS caps & fields support
RDMA/mlx5: Separate DCI QP creation logic
RDMA/mlx5: Add DCS offload support
drivers/infiniband/hw/mlx5/main.c | 10 ++ drivers/infiniband/hw/mlx5/qp.c | 168 ++++++++++++++++++++++++++++++ include/linux/mlx5/mlx5_ifc.h | 14 ++- include/uapi/rdma/mlx5-abi.h | 17 ++- 4 files changed, 204 insertions(+), 5 deletions(-)
Signature · 1 line
2.31.1
PATCH 1/3 net/mlx5: Add DCS caps & fields support
From: Lior Nahmanson <[email protected]>
This fields will be needed when adding a support for DCS offload
max_dci_stream_channels - maximum DCI stream channels supported per DCI. max_dci_errored_streams - maximum DCI error stream channels supported per DCI before a DCI move to error state.
--- include/linux/mlx5/mlx5_ifc.h | 14 +++++++++++--- 1 file changed, 11 insertions(+), 3 deletions(-)
include/linux/mlx5/mlx5_ifc.h +11 -3
@@ -1655,7 +1655,13 @@ struct mlx5_ifc_cmd_hca_cap_bits {u8 max_geneve_tlv_option_data_len[0x5 ];u8 reserved_at_570[0x10 ]; -u8 reserved_at_580[0x33 ]; +u8 reserved_at_580[0xb ]; +u8 log_max_dci_stream_channels[0x5 ]; +u8 reserved_at_590[0x3 ]; +u8 log_max_dci_errored_streams[0x5 ]; +u8 reserved_at_598[0x8 ]; + +u8 reserved_at_5a0[0x13 ];u8 log_max_dek[0x5 ];u8 reserved_at_5b8[0x4 ];u8 mini_cqe_resp_stride_index[0x1 ];
@@ -3013,10 +3019,12 @@ struct mlx5_ifc_qpc_bits {u8 reserved_at_3c0[0x8 ];u8 next_send_psn[0x18 ]; -u8 reserved_at_3e0[0x8 ]; +u8 reserved_at_3e0[0x3 ]; +u8 log_num_dci_stream_channels[0x5 ];u8 cqn_snd[0x18 ]; -u8 reserved_at_400[0x8 ]; +u8 reserved_at_400[0x3 ]; +u8 log_num_dci_errored_streams[0x5 ];u8 deth_sqpn[0x18 ];u8 reserved_at_420[0x20 ];
Signature · 1 line
2.31.1
PATCH 2/3 RDMA/mlx5: Separate DCI QP creation logic
In reply to Leon RomanovskyFrom: Lior Nahmanson <[email protected]>
This patch isolates DCI QP creation logic to separate function, so this change will reduce complexity when adding new features to DCI QP without interfering with other QP types.
The code was copied from create_user_qp() while taking only DCI relevant bits.
--- drivers/infiniband/hw/mlx5/qp.c | 157 ++++++++++++++++++++++++++++++++ 1 file changed, 157 insertions(+)
drivers/infiniband/hw/mlx5/qp.c +157 -0
@@ -1974,6 +1974,160 @@ static int create_xrc_tgt_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,return 0 ; } +static int create_dci (struct mlx5_ib_dev *dev,struct ib_pd *pd, +struct mlx5_ib_qp *qp, +struct mlx5_create_qp_params *params) +{ +struct ib_qp_init_attr *init_attr = params->attr; +struct mlx5_ib_create_qp *ucmd = params->ucmd; +u32 out[MLX5_ST_SZ_DW (create_qp_out)] = {}; +struct ib_udata *udata = params->udata; +u32 uidx = params->uidx; +struct mlx5_ib_resources *devr = &dev->devr; +int inlen =MLX5_ST_SZ_BYTES (create_qp_in); +struct mlx5_core_dev *mdev = dev->mdev; +struct mlx5_ib_cq *send_cq; +struct mlx5_ib_cq *recv_cq; +unsigned long flags; +struct mlx5_ib_qp_base *base; +int ts_format; +int mlx5_st; +void *qpc; +u32 *in; +int err; + +spin_lock_init (&qp->sq.lock); +spin_lock_init (&qp->rq.lock); + + mlx5_st =to_mlx5_st (qp->type); +if (mlx5_st <0 ) +return -EINVAL ; + +if (init_attr->sq_sig_type ==IB_SIGNAL_ALL_WR ) + qp->sq_signal_bits =MLX5_WQE_CTRL_CQ_UPDATE ; + + base = &qp->trans_qp.base; + + qp->has_rq =qp_has_rq (init_attr); + err =set_rq_size (dev, &init_attr->cap, qp->has_rq, qp, ucmd); +if (err) { +mlx5_ib_dbg (dev,"err %d\n" , err); +return err; + } + +if (ucmd->rq_wqe_shift != qp->rq.wqe_shift || + ucmd->rq_wqe_count != qp->rq.wqe_cnt) +return -EINVAL ; + +if (ucmd->sq_wqe_count > (1 <<MLX5_CAP_GEN (mdev, log_max_qp_sz))) +return -EINVAL ; + + ts_format =get_qp_ts_format (dev,to_mcq (init_attr->send_cq), +to_mcq (init_attr->recv_cq)); + +if (ts_format <0 ) +return ts_format; + + err =_create_user_qp (dev, pd, qp, udata, init_attr, &in, ¶ms->resp, + &inlen, base, ucmd); +if (err) +return err; + +if (MLX5_CAP_GEN (mdev, ece_support)) +MLX5_SET (create_qp_in, in, ece, ucmd->ece_options); + qpc =MLX5_ADDR_OF (create_qp_in, in, qpc); + +MLX5_SET (qpc, qpc, st, mlx5_st); +MLX5_SET (qpc, qpc, pm_state,MLX5_QP_PM_MIGRATED ); +MLX5_SET (qpc, qpc, pd,to_mpd (pd)->pdn); + +if (qp->flags_en &MLX5_QP_FLAG_SIGNATURE ) +MLX5_SET (qpc, qpc, wq_signature,1 ); + +if (qp->flags &IB_QP_CREATE_CROSS_CHANNEL ) +MLX5_SET (qpc, qpc, cd_master,1 ); +if (qp->flags &IB_QP_CREATE_MANAGED_SEND ) +MLX5_SET (qpc, qpc, cd_slave_send,1 ); +if (qp->flags_en &MLX5_QP_FLAG_SCATTER_CQE ) +configure_requester_scat_cqe (dev, qp, init_attr, qpc); + +if (qp->rq.wqe_cnt) { +MLX5_SET (qpc, qpc, log_rq_stride, qp->rq.wqe_shift -4 ); +MLX5_SET (qpc, qpc, log_rq_size,ilog2 (qp->rq.wqe_cnt)); + } + +MLX5_SET (qpc, qpc, ts_format, ts_format); +MLX5_SET (qpc, qpc, rq_type,get_rx_type (qp, init_attr)); + +MLX5_SET (qpc, qpc, log_sq_size,ilog2 (qp->sq.wqe_cnt)); + +/* Set default resources */ +if (init_attr->srq) { +MLX5_SET (qpc, qpc, xrcd, devr->xrcdn0); +MLX5_SET (qpc, qpc, srqn_rmpn_xrqn, +to_msrq (init_attr->srq)->msrq.srqn); + }else { +MLX5_SET (qpc, qpc, xrcd, devr->xrcdn1); +MLX5_SET (qpc, qpc, srqn_rmpn_xrqn, +to_msrq (devr->s1)->msrq.srqn); + } + +if (init_attr->send_cq) +MLX5_SET (qpc, qpc, cqn_snd, +to_mcq (init_attr->send_cq)->mcq.cqn); + +if (init_attr->recv_cq) +MLX5_SET (qpc, qpc, cqn_rcv, +to_mcq (init_attr->recv_cq)->mcq.cqn); + +MLX5_SET64 (qpc, qpc, dbr_addr, qp->db.dma); + +/* 0xffffff means we ask to work with cqe version 0 */ +if (MLX5_CAP_GEN (mdev, cqe_version) ==MLX5_CQE_VERSION_V1 ) +MLX5_SET (qpc, qpc, user_index, uidx); + +if (qp->flags &IB_QP_CREATE_PCI_WRITE_END_PADDING ) { +MLX5_SET (qpc, qpc, end_padding_mode, +MLX5_WQ_END_PAD_MODE_ALIGN ); +/* Special case to clean flag */ + qp->flags &= ~IB_QP_CREATE_PCI_WRITE_END_PADDING ; + } + + err =mlx5_qpc_create_qp (dev, &base->mqp, in, inlen, out); + +kvfree (in); +if (err) + goto err_create; + + base->container_mibqp = qp; + base->mqp.event = mlx5_ib_qp_event; +if (MLX5_CAP_GEN (mdev, ece_support)) + params->resp.ece_options =MLX5_GET (create_qp_out, out, ece); + +get_cqs (qp->type, init_attr->send_cq, init_attr->recv_cq, + &send_cq, &recv_cq); +spin_lock_irqsave (&dev->reset_flow_resource_lock, flags); +mlx5_ib_lock_cqs (send_cq, recv_cq); +/* Maintain device to QPs access, needed for further handling via reset +* flow +*/ +list_add_tail (&qp->qps_list, &dev->qp_list); +/* Maintain CQ to QPs access, needed for further handling via reset flow +*/ +if (send_cq) +list_add_tail (&qp->cq_send_list, &send_cq->list_send_qp); +if (recv_cq) +list_add_tail (&qp->cq_recv_list, &recv_cq->list_recv_qp); +mlx5_ib_unlock_cqs (send_cq, recv_cq); +spin_unlock_irqrestore (&dev->reset_flow_resource_lock, flags); + +return 0 ; + +err_create: +destroy_qp (dev, qp, base, udata); +return err; +} +static int create_user_qp (struct mlx5_ib_dev *dev,struct ib_pd *pd,struct mlx5_ib_qp *qp,struct mlx5_create_qp_params *params)
@@ -2840,6 +2994,9 @@ static int create_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,case MLX5_IB_QPT_DCT : err =create_dct (dev, pd, qp, params);break ; +case MLX5_IB_QPT_DCI : + err =create_dci (dev, pd, qp, params); +break ;case IB_QPT_XRC_TGT : err =create_xrc_tgt_qp (dev, qp, params);break ;
Signature · 1 line
2.31.1
Jason Gunthorpe
maintainerRe: [PATCH rdma-next v1 2/3] RDMA/mlx5: Separate DCI QP creation logic
Quoted 18 lines · Leon Romanovsky
From: Lior Nahmanson <[email protected]>
This patch isolates DCI QP creation logic to separate function, so this change will reduce complexity when adding new features to DCI QP without interfering with other QP types.
The code was copied from
create_user_qp()while taking only DCI relevant bits.Reviewed-by Meir Lichtinger <[email protected]>Signed-off-by Lior Nahmanson <[email protected]>Signed-off-by Leon Romanovsky <[email protected]>drivers/infiniband/hw/mlx5/qp.c | 157 ++++++++++++++++++++++++++++++++ 1 file changed, 157 insertions(+)drivers/infiniband/hw/mlx5/qp.c +1
-0+++ b/drivers/infiniband/hw/mlx5/qp.c
drivers/infiniband/hw/mlx5/qp.c static int create_xrc_tgt_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp, · line 1980
Quoted 7 lines
drivers/infiniband/hw/mlx5/qp.c +4
-0@@ -1974,6 +1974,160 @@ static int create_xrc_tgt_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,return 0 ; } +static int create_dci (struct mlx5_ib_dev *dev,struct ib_pd *pd, +struct mlx5_ib_qp *qp, +struct mlx5_create_qp_params *params) +{
This is a huge amount of copying just to add 4 lines, why?
There must be a better way to do this qp stuff.
Why not put more stuff in _create_user_qp()?
Jason
Re: [PATCH rdma-next v1 2/3] RDMA/mlx5: Separate DCI QP creation logic
Quoted 29 lines · Jason Gunthorpe
Quoted 18 lines · Leon Romanovsky
From: Lior Nahmanson <[email protected]>
This patch isolates DCI QP creation logic to separate function, so this change will reduce complexity when adding new features to DCI QP without interfering with other QP types.
The code was copied from
create_user_qp()while taking only DCI relevant bits.Reviewed-by Meir Lichtinger <[email protected]>Signed-off-by Lior Nahmanson <[email protected]>Signed-off-by Leon Romanovsky <[email protected]>drivers/infiniband/hw/mlx5/qp.c | 157 ++++++++++++++++++++++++++++++++ 1 file changed, 157 insertions(+)drivers/infiniband/hw/mlx5/qp.c +1
-0+++ b/drivers/infiniband/hw/mlx5/qp.cdrivers/infiniband/hw/mlx5/qp.c static int create_xrc_tgt_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp, · line 1980
Quoted 7 lines
drivers/infiniband/hw/mlx5/qp.c +4
-0@@ -1974,6 +1974,160 @@ static int create_xrc_tgt_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,return 0 ; } +static int create_dci (struct mlx5_ib_dev *dev,struct ib_pd *pd, +struct mlx5_ib_qp *qp, +struct mlx5_create_qp_params *params) +{This is a huge amount of copying just to add 4 lines, why?
Quoted 3 lines
There must be a better way to do this qp stuff.
Why not put more stuff in
_create_user_qp()?
Lior proposed it in original patch, but I didn't like it. It caused to mix of various QP types and maze of "if () else ()" that are not applicable one to another.
The huge _create_user_qp() is the reason why create_dci() is not small, we simply had hard time to understand if specific HW bit is needed or not in DCI flow.
My goal is to have small per-QP type specific functions that calls to simple functions for very narrow scope.
Something like that:
static int create_dci(...)
{
...
configure_send_cq(..)
configure_recv_sq(..)
configure_srq(...)
...
}
static int create_user_qp(...)
{
...
configure_send_cq(..)
configure_recv_sq(..)
configure_srq(...)
...
}
Thanks
Quoted 2 lines
Jason
PATCH 3/3 RDMA/mlx5: Add DCS offload support
In reply to Leon RomanovskyFrom: Lior Nahmanson <[email protected]>
DCS is an offload to SW load balancing of DC initiator work requests.
A single DCI can be connected to only one target at the time and can't start new connection until the previous work request is completed. This limitation will cause to delay when the initiator process needs to transfer data to multiple targets at the same time.
The SW solution is to use a process that handling and spreading the work request on many DCIs according to destinations.
This feature is an offload to this process and coming to reduce the load from the CPU and improve the performance.
--- drivers/infiniband/hw/mlx5/main.c | 10 ++++++++++ drivers/infiniband/hw/mlx5/qp.c | 11 +++++++++++ include/uapi/rdma/mlx5-abi.h | 17 +++++++++++++++-- 3 files changed, 36 insertions(+), 2 deletions(-)
drivers/infiniband/hw/mlx5/main.c +10 -0
@@ -1174,6 +1174,16 @@ static int mlx5_ib_query_device(struct ib_device *ibdev,MLX5_IB_TUNNELED_OFFLOADS_MPLS_UDP ; } +if (offsetofend (typeof (resp), dci_streams_caps) <= uhw_outlen) { + resp.response_length +=sizeof (resp.dci_streams_caps); + + resp.dci_streams_caps.max_log_num_concurent = +MLX5_CAP_GEN (mdev, log_max_dci_stream_channels); + + resp.dci_streams_caps.max_log_num_errored = +MLX5_CAP_GEN (mdev, log_max_dci_errored_streams); + } +if (uhw_outlen) { err =ib_copy_to_udata (uhw, &resp, resp.response_length);
drivers/infiniband/hw/mlx5/qp.c +11 -0
@@ -2056,6 +2056,13 @@ static int create_dci(struct mlx5_ib_dev *dev, struct ib_pd *pd,MLX5_SET (qpc, qpc, log_rq_size,ilog2 (qp->rq.wqe_cnt)); } +if (qp->flags_en &MLX5_QP_FLAG_DCI_STREAM ) { +MLX5_SET (qpc, qpc, log_num_dci_stream_channels, + ucmd->dci_streams.log_num_concurent); +MLX5_SET (qpc, qpc, log_num_dci_errored_streams, + ucmd->dci_streams.log_num_errored); + } +MLX5_SET (qpc, qpc, ts_format, ts_format);MLX5_SET (qpc, qpc, rq_type,get_rx_type (qp, init_attr));
@@ -2799,6 +2806,10 @@ static int process_vendor_flags(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,process_vendor_flag (dev, &flags,MLX5_QP_FLAG_TYPE_DCI , true, qp);process_vendor_flag (dev, &flags,MLX5_QP_FLAG_TYPE_DCT , true, qp); +process_vendor_flag (dev, &flags,MLX5_QP_FLAG_DCI_STREAM , +MLX5_CAP_GEN (mdev, log_max_dci_stream_channels) && +MLX5_CAP_GEN (mdev, log_max_dci_errored_streams), + qp);process_vendor_flag (dev, &flags,MLX5_QP_FLAG_SIGNATURE , true, qp); process_vendor_flag(dev, &flags, MLX5_QP_FLAG_SCATTER_CQE,
include/uapi/rdma/mlx5-abi.h +15 -2
@@ -50,6 +50,7 @@ enum {MLX5_QP_FLAG_ALLOW_SCATTER_CQE =1 <<8 ,MLX5_QP_FLAG_PACKET_BASED_CREDIT_MODE =1 <<9 ,MLX5_QP_FLAG_UAR_PAGE_INDEX =1 <<10 , + MLX5_QP_FLAG_DCI_STREAM = 1 << 11, }; enum {
@@ -237,6 +238,11 @@ struct mlx5_ib_striding_rq_caps {__u32 reserved; }; +struct mlx5_ib_dci_streams_caps { +__u8 max_log_num_concurent; +__u8 max_log_num_errored; +}; +enum mlx5_ib_query_dev_resp_flags {/* Support 128B CQE compression */ MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_COMP =1 <<0 ,
@@ -265,7 +271,8 @@ struct mlx5_ib_query_device_resp {struct mlx5_ib_sw_parsing_caps sw_parsing_caps;struct mlx5_ib_striding_rq_caps striding_rq_caps;__u32 tunnel_offloads_caps;/* enum mlx5_ib_tunnel_offloads */ -__u32 reserved; +struct mlx5_ib_dci_streams_caps dci_streams_caps; +__u16 reserved; };enum mlx5_ib_create_cq_flags {
@@ -311,6 +318,11 @@ struct mlx5_ib_create_srq_resp {__u32 reserved; }; +struct mlx5_ib_create_qp_dci_streams { +__u8 log_num_concurent; +__u8 log_num_errored; +}; +struct mlx5_ib_create_qp {__aligned_u64 buf_addr;__aligned_u64 db_addr;
@@ -325,7 +337,8 @@ struct mlx5_ib_create_qp {__aligned_u64 access_key; };__u32 ece_options; -__u32 reserved; +struct mlx5_ib_create_qp_dci_streams dci_streams; +__u16 reserved; };/* RX Hash function flags */
Signature · 1 line
2.31.1
Jason Gunthorpe
maintainerQuoted 18 lines · Leon Romanovsky
From: Leon Romanovsky <[email protected]>
Changelog:
v1:
* Rephrase commit message of second patch
v0: https://lore.kernel.org/linux-rdma/[email protected]This patchset from Lior adds support of DCI stream channel (DCS) support.
DCS is an offload to SW load balancing of DC initiator work requests.
A single DC QP initiator (DCI) can be connected to only one target at the time and can't start new connection until the previous work request is completed.
This limitation causes to delays when the initiator process needs to transfer data to multiple targets at the same time.
Quoted 6 lines
Thanks
Lior Nahmanson (3):
net/mlx5: Add DCS caps & fields support
RDMA/mlx5: Separate DCI QP creation logic
RDMA/mlx5: Add DCS offload support
Okay, can you update the shared branch?
Thanks,
Jason
Quoted 19 lines · Jason Gunthorpe
Quoted 18 lines · Leon Romanovsky
From: Leon Romanovsky <[email protected]>
Changelog:
v1:
* Rephrase commit message of second patch
v0: https://lore.kernel.org/linux-rdma/[email protected]This patchset from Lior adds support of DCI stream channel (DCS) support.
DCS is an offload to SW load balancing of DC initiator work requests.
A single DC QP initiator (DCI) can be connected to only one target at the time and can't start new connection until the previous work request is completed.
This limitation causes to delays when the initiator process needs to transfer data to multiple targets at the same time.
Quoted 8 lines
Quoted 6 lines
Thanks
Lior Nahmanson (3):
net/mlx5: Add DCS caps & fields support
RDMA/mlx5: Separate DCI QP creation logic
RDMA/mlx5: Add DCS offload supportOkay, can you update the shared branch?
Pushed, 96cd2dd65bb0 ("net/mlx5: Add DCS caps & fields support")
Thanks
Quoted 3 lines
Thanks,
Jason
Quoted 19 lines · Leon Romanovsky
From: Leon Romanovsky <[email protected]>
Changelog:
v1:
* Rephrase commit message of second patch
v0: https://lore.kernel.org/linux-rdma/[email protected]------------
This patchset from Lior adds support of DCI stream channel (DCS) support.
DCS is an offload to SW load balancing of DC initiator work requests.
A single DC QP initiator (DCI) can be connected to only one target at the time and can't start new connection until the previous work request is completed.
This limitation causes to delays when the initiator process needs to transfer data to multiple targets at the same time.
Quoted 6 lines
Thanks
Lior Nahmanson (3):
net/mlx5: Add DCS caps & fields support
RDMA/mlx5: Separate DCI QP creation logic
RDMA/mlx5: Add DCS offload support
Applied to for-next, thanks
Jason
Quoted 20 lines · Jason Gunthorpe
Quoted 19 lines · Leon Romanovsky
From: Leon Romanovsky <[email protected]>
Changelog:
v1:
* Rephrase commit message of second patch
v0: https://lore.kernel.org/linux-rdma/[email protected]------------
This patchset from Lior adds support of DCI stream channel (DCS) support.
DCS is an offload to SW load balancing of DC initiator work requests.
A single DC QP initiator (DCI) can be connected to only one target at the time and can't start new connection until the previous work request is completed.
This limitation causes to delays when the initiator process needs to transfer data to multiple targets at the same time.
Quoted 8 lines
Quoted 6 lines
Thanks
Lior Nahmanson (3):
net/mlx5: Add DCS caps & fields support
RDMA/mlx5: Separate DCI QP creation logic
RDMA/mlx5: Add DCS offload supportApplied to for-next, thanks
Jason,
You forgot to push this to the RDMA tree.
Thanks
Quoted 2 lines
Jason