Driver part of the ECE
Subsystems (5)
v2 cover Driver part of the ECE
From: Leon Romanovsky <[email protected]>
Changelog:
v2:
* Rebased on latest wip/jgg-rdma-next branch, commit a94dae867c56
* Fixed wrong setting of pm_state field in mlx5 conversion patch
* Removed DC support for now
v1: https://lore.kernel.org/linux-rdma/[email protected]
* Fixed compatibility issue of "old" kernel vs. "new" rdma-core. This is handled in extra patch.
* Improved comments and commit messages after feedback from Yishai.
* Added Mark Z. ROB tags
v0: https://lore.kernel.org/linux-rdma/[email protected]
----------------------------------------------------------------------
Hi,
This is driver part of the RDMA-CM ECE series [1].
According to the IBTA, ECE data is completely vendor specific, so this series extends mlx5_ib create_qp and modify_qp structs with extra field to pass ECE options to/from the application.
Thanks
[1]
https://lore.kernel.org/linux-rdma/[email protected]
Leon Romanovsky (9):
net/mlx5: Add ability to read and write ECE options RDMA/mlx5: Get ECE options from FW during create QP RDMA/mlx5: Set ECE options during QP create RDMA/mlx5: Use direct modify QP implementation RDMA/mlx5: Remove manually crafted QP context the query call RDMA/mlx5: Convert modify QP to use MLX5_SET macros RDMA/mlx5: Advertise ECE support
RDMA/mlx5: Set ECE options during modify QP RDMA/mlx5: Return ECE data after modify QP
drivers/infiniband/hw/mlx5/main.c | 3 + drivers/infiniband/hw/mlx5/qp.c | 478 +++++++++++++++++------------- drivers/infiniband/hw/mlx5/qp.h | 6 +- drivers/infiniband/hw/mlx5/qpc.c | 44 ++- include/linux/mlx5/mlx5_ifc.h | 24 +- include/linux/mlx5/qp.h | 66 ----- include/uapi/rdma/mlx5-abi.h | 9 +- 7 files changed, 333 insertions(+), 297 deletions(-)
--
2.26.2
PATCH v2 2/9 RDMA/mlx5: Get ECE options from FW during create QP
From: Leon Romanovsky <[email protected]>
Supported ECE options are returned from FW in the create_qp phase and zero means that field is not valid. Such default value allows us to reuse reserved field without worries about comp_mask.
Update create QP API to return ECE options.
--- drivers/infiniband/hw/mlx5/qp.c | 16 +++++++++++----- drivers/infiniband/hw/mlx5/qp.h | 4 ++-- drivers/infiniband/hw/mlx5/qpc.c | 8 ++++---- include/uapi/rdma/mlx5-abi.h | 2 +- 4 files changed, 18 insertions(+), 12 deletions(-)
drivers/infiniband/hw/mlx5/qp.c +11 -5
@@ -1842,6 +1842,7 @@ static int create_xrc_tgt_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,struct ib_qp_init_attr *attr = params->attr;u32 uidx = params->uidx;struct mlx5_ib_resources *devr = &dev->devr; +u32 out[MLX5_ST_SZ_DW (create_qp_out)] = {};int inlen =MLX5_ST_SZ_BYTES (create_qp_in);struct mlx5_core_dev *mdev = dev->mdev;struct mlx5_ib_qp_base *base;
@@ -1894,13 +1895,14 @@ static int create_xrc_tgt_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp, } base = &qp->trans_qp.base; - err =mlx5_core_create_qp (dev, &base->mqp, in, inlen); + err =mlx5_qpc_create_qp (dev, &base->mqp, in, inlen, out);kvfree (in);if (err)return err; base->container_mibqp = qp; base->mqp.event = mlx5_ib_qp_event; + params->resp.ece_options =MLX5_GET (create_qp_out, out, ece);spin_lock_irqsave (&dev->reset_flow_resource_lock, flags);list_add_tail (&qp->qps_list, &dev->qp_list);
@@ -1916,6 +1918,7 @@ static int create_user_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd, {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;
@@ -2065,7 +2068,7 @@ static int create_user_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd, err =create_raw_packet_qp (dev, qp, in, inlen, pd, udata, ¶ms->resp); }else - err =mlx5_core_create_qp (dev, &base->mqp, in, inlen); + err =mlx5_qpc_create_qp (dev, &base->mqp, in, inlen, out);kvfree (in);if (err)
@@ -2073,6 +2076,7 @@ static int create_user_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd, base->container_mibqp = qp; base->mqp.event = mlx5_ib_qp_event; + 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);
@@ -2105,6 +2109,7 @@ static int create_kernel_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,struct ib_qp_init_attr *attr = params->attr;u32 uidx = params->uidx;struct mlx5_ib_resources *devr = &dev->devr; +u32 out[MLX5_ST_SZ_DW (create_qp_out)] = {};int inlen =MLX5_ST_SZ_BYTES (create_qp_in);struct mlx5_core_dev *mdev = dev->mdev;struct mlx5_ib_cq *send_cq;
@@ -2195,7 +2200,7 @@ static int create_kernel_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,if (qp->flags &IB_QP_CREATE_IPOIB_UD_LSO )MLX5_SET (qpc, qpc, ulp_stateless_offload_mode,1 ); - err =mlx5_core_create_qp (dev, &base->mqp, in, inlen); + err =mlx5_qpc_create_qp (dev, &base->mqp, in, inlen, out);kvfree (in);if (err) goto err_create;
@@ -2779,12 +2784,13 @@ static int create_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd, qp->ibqp.qp_num = qp->trans_qp.base.mqp.qpn;mlx5_ib_dbg (dev, -"QP type %d, ib qpn 0x%X, mlx qpn 0x%x, rcqn 0x%x, scqn 0x%x\n" , +"QP type %d, ib qpn 0x%X, mlx qpn 0x%x, rcqn 0x%x, scqn 0x%x, ece 0x%x\n" , qp->type, qp->ibqp.qp_num, qp->trans_qp.base.mqp.qpn, params->attr->recv_cq ?to_mcq (params->attr->recv_cq)->mcq.cqn :-1 , params->attr->send_cq ?to_mcq (params->attr->send_cq)->mcq.cqn : --1 ); +-1 , + params->resp.ece_options);return 0 ; }
drivers/infiniband/hw/mlx5/qp.h +2 -2
@@ -13,8 +13,8 @@ void mlx5_cleanup_qp_table(struct mlx5_ib_dev *dev);int mlx5_core_create_dct (struct mlx5_ib_dev *dev,struct mlx5_core_dct *qp,u32 *in,int inlen,u32 *out,int outlen); -int mlx5_core_create_qp (struct mlx5_ib_dev *dev,struct mlx5_core_qp *qp, -u32 *in,int inlen); +int mlx5_qpc_create_qp (struct mlx5_ib_dev *dev,struct mlx5_core_qp *qp, +u32 *in,int inlen,u32 *out);int mlx5_core_qp_modify (struct mlx5_ib_dev *dev,u16 opcode,u32 opt_param_mask,void *qpc,struct mlx5_core_qp *qp);int mlx5_core_destroy_qp (struct mlx5_ib_dev *dev,struct mlx5_core_qp *qp);
drivers/infiniband/hw/mlx5/qpc.c +4 -4
@@ -236,16 +236,16 @@ int mlx5_core_create_dct(struct mlx5_ib_dev *dev, struct mlx5_core_dct *dct,return err; } -int mlx5_core_create_qp (struct mlx5_ib_dev *dev,struct mlx5_core_qp *qp, -u32 *in,int inlen) +int mlx5_qpc_create_qp (struct mlx5_ib_dev *dev,struct mlx5_core_qp *qp, +u32 *in,int inlen,u32 *out) { -u32 out[MLX5_ST_SZ_DW (create_qp_out)] = {};u32 din[MLX5_ST_SZ_DW (destroy_qp_in)] = {};int err;MLX5_SET (create_qp_in, in, opcode,MLX5_CMD_OP_CREATE_QP ); - err =mlx5_cmd_exec (dev->mdev, in, inlen, out,sizeof (out)); + err =mlx5_cmd_exec (dev->mdev, in, inlen, out, +MLX5_ST_SZ_BYTES (create_qp_out));if (err)return err;
include/uapi/rdma/mlx5-abi.h +1 -1
@@ -371,7 +371,7 @@ enum mlx5_ib_create_qp_resp_mask {struct mlx5_ib_create_qp_resp {__u32 bfreg_index; -__u32 reserved; +__u32 ece_options;__u32 comp_mask;__u32 tirn;__u32 tisn;
Signature · 1 line
2.26.2
PATCH v2 1/9 net/mlx5: Add ability to read and write ECE options
In reply to Leon RomanovskyFrom: Leon Romanovsky <[email protected]>
The end result of RDMA-CM ECE handshake is ECE options, which is needed to be used while configuring data QPs. Such options can come in any QP state, so add in/out fields to set and query ECE options.
OUT fields:
* create_qp() - default ECE options for that type of QP.
* modify_qp() - enabled ECE options after QP state transition.
IN fields:
* create_qp() - create QP with this ECE option.
* modify_qp() - requested options. For unconnected QPs, the FW will return an error if ECE is already configured with any options that not equal to previously set.
--- include/linux/mlx5/mlx5_ifc.h | 24 ++++++++++++++---------- 1 file changed, 14 insertions(+), 10 deletions(-)
include/linux/mlx5/mlx5_ifc.h +14 -10
@@ -1296,7 +1296,7 @@ struct mlx5_ifc_cmd_hca_cap_bits {u8 port_type[0x2 ];u8 num_ports[0x8 ]; -u8 reserved_at_1c0[0x1 ]; +u8 ece_support[0x1 ];u8 pps[0x1 ];u8 pps_modify[0x1 ];u8 log_max_msg[0x5 ];
@@ -4220,7 +4220,8 @@ struct mlx5_ifc_rts2rts_qp_out_bits {u8 syndrome[0x20 ]; -u8 reserved_at_40[0x40 ]; +u8 reserved_at_40[0x20 ]; +u8 ece[0x20 ]; };struct mlx5_ifc_rts2rts_qp_in_bits {
@@ -4237,7 +4238,7 @@ struct mlx5_ifc_rts2rts_qp_in_bits {u8 opt_param_mask[0x20 ]; -u8 reserved_at_a0[0x20 ]; +u8 ece[0x20 ];struct mlx5_ifc_qpc_bits qpc;
@@ -4250,7 +4251,8 @@ struct mlx5_ifc_rtr2rts_qp_out_bits {u8 syndrome[0x20 ]; -u8 reserved_at_40[0x40 ]; +u8 reserved_at_40[0x20 ]; +u8 ece[0x20 ]; };struct mlx5_ifc_rtr2rts_qp_in_bits {
@@ -4267,7 +4269,7 @@ struct mlx5_ifc_rtr2rts_qp_in_bits {u8 opt_param_mask[0x20 ]; -u8 reserved_at_a0[0x20 ]; +u8 ece[0x20 ];struct mlx5_ifc_qpc_bits qpc;
@@ -4819,7 +4821,8 @@ struct mlx5_ifc_query_qp_out_bits {u8 syndrome[0x20 ]; -u8 reserved_at_40[0x40 ]; +u8 reserved_at_40[0x20 ]; +u8 ece[0x20 ];u8 opt_param_mask[0x20 ];
@@ -6584,7 +6587,8 @@ struct mlx5_ifc_init2rtr_qp_out_bits {u8 syndrome[0x20 ]; -u8 reserved_at_40[0x40 ]; +u8 reserved_at_40[0x20 ]; +u8 ece[0x20 ]; };struct mlx5_ifc_init2rtr_qp_in_bits {
@@ -6601,7 +6605,7 @@ struct mlx5_ifc_init2rtr_qp_in_bits {u8 opt_param_mask[0x20 ]; -u8 reserved_at_a0[0x20 ]; +u8 ece[0x20 ];struct mlx5_ifc_qpc_bits qpc;
@@ -7697,7 +7701,7 @@ struct mlx5_ifc_create_qp_out_bits {u8 reserved_at_40[0x8 ];u8 qpn[0x18 ]; -u8 reserved_at_60[0x20 ]; +u8 ece[0x20 ]; };struct mlx5_ifc_create_qp_in_bits {
@@ -7711,7 +7715,7 @@ struct mlx5_ifc_create_qp_in_bits {u8 opt_param_mask[0x20 ]; -u8 reserved_at_a0[0x20 ]; +u8 ece[0x20 ];struct mlx5_ifc_qpc_bits qpc;
Signature · 1 line
2.26.2
PATCH v2 4/9 RDMA/mlx5: Use direct modify QP implementation
In reply to Leon RomanovskyFrom: Leon Romanovsky <[email protected]>
As a preparation to removal hand crafted mlx5_qp_context, convert counter code to use mlx5_cmd_exec_in() directly.
--- drivers/infiniband/hw/mlx5/qp.c | 17 +++++++++++------ 1 file changed, 11 insertions(+), 6 deletions(-)
drivers/infiniband/hw/mlx5/qp.c +11 -6
@@ -3688,10 +3688,11 @@ static int __mlx5_ib_qp_set_counter(struct ib_qp *qp, struct rdma_counter *counter) {struct mlx5_ib_dev *dev =to_mdev (qp->device); +u32 in[MLX5_ST_SZ_DW (rts2rts_qp_in)] = {};struct mlx5_ib_qp *mqp =to_mqp (qp); -struct mlx5_qp_context context = {};struct mlx5_ib_qp_base *base;u32 set_id; +u32 *qpc;if (counter) set_id = counter->id;
@@ -3699,11 +3700,15 @@ static int __mlx5_ib_qp_set_counter(struct ib_qp *qp, set_id =mlx5_ib_get_counters_id (dev, mqp->port -1 ); base = &mqp->trans_qp.base; - context.qp_counter_set_usr_page &=cpu_to_be32 (0xffffff ); - context.qp_counter_set_usr_page |=cpu_to_be32 (set_id <<24 ); -return mlx5_core_qp_modify (dev,MLX5_CMD_OP_RTS2RTS_QP , -MLX5_QP_OPTPAR_COUNTER_SET_ID , &context, - &base->mqp); +MLX5_SET (rts2rts_qp_in, in, opcode,MLX5_CMD_OP_RTS2RTS_QP ); +MLX5_SET (rts2rts_qp_in, in, qpn, base->mqp.qpn); +MLX5_SET (rts2rts_qp_in, in, uid, base->mqp.uid); +MLX5_SET (rts2rts_qp_in, in, opt_param_mask, +MLX5_QP_OPTPAR_COUNTER_SET_ID ); + + qpc =MLX5_ADDR_OF (rts2rts_qp_in, in, qpc); +MLX5_SET (qpc, qpc, counter_set_id, set_id); +return mlx5_cmd_exec_in (dev->mdev, rts2rts_qp, in); } static int __mlx5_ib_modify_qp(struct ib_qp *ibqp,
Signature · 1 line
2.26.2
PATCH v2 5/9 RDMA/mlx5: Remove manually crafted QP context the query call
In reply to Leon RomanovskyFrom: Leon Romanovsky <[email protected]>
As a preparation to removal hand crafted mlx5_qp_context, convert query_qp_attr() to use proper MLX5_GET() macros.
--- drivers/infiniband/hw/mlx5/qp.c | 129 ++++++++++++++------------------ 1 file changed, 56 insertions(+), 73 deletions(-)
drivers/infiniband/hw/mlx5/qp.c +56 -73
@@ -4330,50 +4330,35 @@ static inline enum ib_mig_state to_ib_mig_state(int mlx5_mig_state) } } -static int to_ib_qp_access_flags (int mlx5_flags) -{ -int ib_flags =0 ; - -if (mlx5_flags &MLX5_QP_BIT_RRE ) - ib_flags |=IB_ACCESS_REMOTE_READ ; -if (mlx5_flags &MLX5_QP_BIT_RWE ) - ib_flags |=IB_ACCESS_REMOTE_WRITE ; -if (mlx5_flags &MLX5_QP_BIT_RAE ) - ib_flags |=IB_ACCESS_REMOTE_ATOMIC ; - -return ib_flags; -} -static void to_rdma_ah_attr (struct mlx5_ib_dev *ibdev, -struct rdma_ah_attr *ah_attr, -struct mlx5_qp_path *path) +struct rdma_ah_attr *ah_attr,void *path) { +int port =MLX5_GET (ads, path, vhca_port_num); +int static_rate;memset (ah_attr,0 ,sizeof (*ah_attr)); -if (!path->port || path->port > ibdev->num_ports) +if (!port || port > ibdev->num_ports)return ; - ah_attr->type =rdma_ah_find_type (&ibdev->ib_dev, path->port); + ah_attr->type =rdma_ah_find_type (&ibdev->ib_dev, port); -rdma_ah_set_port_num (ah_attr, path->port); -rdma_ah_set_sl (ah_attr, path->dci_cfi_prio_sl &0xf ); +rdma_ah_set_port_num (ah_attr, port); +rdma_ah_set_sl (ah_attr,MLX5_GET (ads, path, sl)); -rdma_ah_set_dlid (ah_attr,be16_to_cpu (path->rlid)); -rdma_ah_set_path_bits (ah_attr, path->grh_mlid &0x7f ); -rdma_ah_set_static_rate (ah_attr, - path->static_rate ? path->static_rate -5 :0 ); +rdma_ah_set_dlid (ah_attr,MLX5_GET (ads, path, rlid)); +rdma_ah_set_path_bits (ah_attr,MLX5_GET (ads, path, mlid)); -if (path->grh_mlid & (1 <<7 ) || + static_rate =MLX5_GET (ads, path, stat_rate); +rdma_ah_set_static_rate (ah_attr, static_rate ? static_rate -5 :0 ); +if (MLX5_GET (ads, path, grh) || ah_attr->type ==RDMA_AH_ATTR_TYPE_ROCE ) { -u32 tc_fl =be32_to_cpu (path->tclass_flowlabel); - -rdma_ah_set_grh (ah_attr,NULL , - tc_fl &0xfffff , - path->mgid_index, - path->hop_limit, - (tc_fl >>20 ) &0xff ); -rdma_ah_set_dgid_raw (ah_attr, path->rgid); +rdma_ah_set_grh (ah_attr,NULL ,MLX5_GET (ads, path, flow_label), +MLX5_GET (ads, path, src_addr_index), +MLX5_GET (ads, path, hop_limit), +MLX5_GET (ads, path, tclass)); +memcpy (ah_attr,MLX5_ADDR_OF (ads, path, rgid_rip), +MLX5_FLD_SZ_BYTES (ads, rgid_rip)); } }
@@ -4495,10 +4480,9 @@ static int query_qp_attr(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp, struct ib_qp_attr *qp_attr) {int outlen =MLX5_ST_SZ_BYTES (query_qp_out); -struct mlx5_qp_context *context; -int mlx5_state; +void *qpc, *pri_path, *alt_path;u32 *outb; -int err =0 ; +int err; outb =kzalloc (outlen,GFP_KERNEL );if (!outb)
@@ -4508,47 +4492,46 @@ static int query_qp_attr(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,if (err) goto out; -/* FIXME: use MLX5_GET rather than mlx5_qp_context manual struct */ - context = (struct mlx5_qp_context *)MLX5_ADDR_OF (query_qp_out, outb, qpc); + qpc =MLX5_ADDR_OF (query_qp_out, outb, qpc); + + qp->state =to_ib_qp_state (MLX5_GET (qpc, qpc, state)); +if (MLX5_GET (qpc, qpc, state) ==MLX5_QP_STATE_SQ_DRAINING ) + qp_attr->sq_draining =1 ; + + qp_attr->path_mtu =MLX5_GET (qpc, qpc, mtu); + qp_attr->path_mig_state =to_ib_mig_state (MLX5_GET (qpc, qpc, pm_state)); + qp_attr->qkey =MLX5_GET (qpc, qpc, q_key); + qp_attr->rq_psn =MLX5_GET (qpc, qpc, next_rcv_psn); + qp_attr->sq_psn =MLX5_GET (qpc, qpc, next_send_psn); + qp_attr->dest_qp_num =MLX5_GET (qpc, qpc, remote_qpn); - mlx5_state =be32_to_cpu (context->flags) >>28 ; +if (MLX5_GET (qpc, qpc, rre)) + qp_attr->qp_access_flags |=IB_ACCESS_REMOTE_READ ; +if (MLX5_GET (qpc, qpc, rwe)) + qp_attr->qp_access_flags |=IB_ACCESS_REMOTE_WRITE ; +if (MLX5_GET (qpc, qpc, rae)) + qp_attr->qp_access_flags |=IB_ACCESS_REMOTE_ATOMIC ; - qp->state =to_ib_qp_state (mlx5_state); - qp_attr->path_mtu = context->mtu_msgmax >>5 ; - qp_attr->path_mig_state = -to_ib_mig_state ((be32_to_cpu (context->flags) >>11 ) &0x3 ); - qp_attr->qkey =be32_to_cpu (context->qkey); - qp_attr->rq_psn =be32_to_cpu (context->rnr_nextrecvpsn) &0xffffff ; - qp_attr->sq_psn =be32_to_cpu (context->next_send_psn) &0xffffff ; - qp_attr->dest_qp_num =be32_to_cpu (context->log_pg_sz_remote_qpn) &0xffffff ; - qp_attr->qp_access_flags = -to_ib_qp_access_flags (be32_to_cpu (context->params2)); + qp_attr->max_rd_atomic =1 <<MLX5_GET (qpc, qpc, log_sra_max); + qp_attr->max_dest_rd_atomic =1 <<MLX5_GET (qpc, qpc, log_rra_max); + qp_attr->min_rnr_timer =MLX5_GET (qpc, qpc, min_rnr_nak); + qp_attr->retry_cnt =MLX5_GET (qpc, qpc, retry_count); + qp_attr->rnr_retry =MLX5_GET (qpc, qpc, rnr_retry); + + pri_path =MLX5_ADDR_OF (qpc, qpc, primary_address_path); + alt_path =MLX5_ADDR_OF (qpc, qpc, secondary_address_path);if (qp->ibqp.qp_type ==IB_QPT_RC || qp->ibqp.qp_type ==IB_QPT_UC ) { -to_rdma_ah_attr (dev, &qp_attr->ah_attr, &context->pri_path); -to_rdma_ah_attr (dev, &qp_attr->alt_ah_attr, &context->alt_path); - qp_attr->alt_pkey_index = -be16_to_cpu (context->alt_path.pkey_index); - qp_attr->alt_port_num = -rdma_ah_get_port_num (&qp_attr->alt_ah_attr); - } - - qp_attr->pkey_index =be16_to_cpu (context->pri_path.pkey_index); - qp_attr->port_num = context->pri_path.port; - -/* qp_attr->en_sqd_async_notify is only applicable in modify qp */ - qp_attr->sq_draining = mlx5_state ==MLX5_QP_STATE_SQ_DRAINING ; - - qp_attr->max_rd_atomic =1 << ((be32_to_cpu (context->params1) >>21 ) &0x7 ); - - qp_attr->max_dest_rd_atomic = -1 << ((be32_to_cpu (context->params2) >>21 ) &0x7 ); - qp_attr->min_rnr_timer = - (be32_to_cpu (context->rnr_nextrecvpsn) >>24 ) &0x1f ; - qp_attr->timeout = context->pri_path.ackto_lt >>3 ; - qp_attr->retry_cnt = (be32_to_cpu (context->params1) >>16 ) &0x7 ; - qp_attr->rnr_retry = (be32_to_cpu (context->params1) >>13 ) &0x7 ; - qp_attr->alt_timeout = context->alt_path.ackto_lt >>3 ; +to_rdma_ah_attr (dev, &qp_attr->ah_attr, pri_path); +to_rdma_ah_attr (dev, &qp_attr->alt_ah_attr, alt_path); + qp_attr->alt_pkey_index =MLX5_GET (ads, alt_path, pkey_index); + qp_attr->alt_port_num =MLX5_GET (ads, alt_path, vhca_port_num); + } + + qp_attr->pkey_index =MLX5_GET (ads, pri_path, pkey_index); + qp_attr->port_num =MLX5_GET (ads, pri_path, vhca_port_num); + qp_attr->timeout =MLX5_GET (ads, pri_path, ack_timeout); + qp_attr->alt_timeout =MLX5_GET (ads, alt_path, ack_timeout); out:kfree (outb);
Signature · 1 line
2.26.2
PATCH v2 3/9 RDMA/mlx5: Set ECE options during QP create
In reply to Leon RomanovskyFrom: Leon Romanovsky <[email protected]>
Allow users to ask creation of QPs with specific ECE options. Such early set even before RDMA-CM connection is established is useful if user knows exactly which option he needs.
--- drivers/infiniband/hw/mlx5/qp.c | 75 +++++++++++++++++++++++++++++---- include/uapi/rdma/mlx5-abi.h | 2 + 2 files changed, 68 insertions(+), 9 deletions(-)
drivers/infiniband/hw/mlx5/qp.c +66 -9
@@ -1552,6 +1552,7 @@ struct mlx5_create_qp_params {struct ib_udata *udata;size_t inlen;size_t outlen; +size_t ucmd_size;void *ucmd; u8 is_rss_raw : 1;struct ib_qp_init_attr *attr;
@@ -1839,6 +1840,7 @@ static int get_atomic_mode(struct mlx5_ib_dev *dev,static int create_xrc_tgt_qp (struct mlx5_ib_dev *dev,struct mlx5_ib_qp *qp,struct mlx5_create_qp_params *params) { +struct mlx5_ib_create_qp *ucmd = params->ucmd;struct ib_qp_init_attr *attr = params->attr;u32 uidx = params->uidx;struct mlx5_ib_resources *devr = &dev->devr;
@@ -1860,6 +1862,8 @@ static int create_xrc_tgt_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,if (!in)return -ENOMEM ; +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_QP_ST_XRC );
@@ -1974,6 +1978,8 @@ static int create_user_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,if (is_sqp (init_attr->qp_type)) qp->port = init_attr->port_num; +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);
@@ -2709,19 +2715,22 @@ static int process_udata_size(struct mlx5_ib_dev *dev, struct mlx5_create_qp_params *params) {size_t ucmd =sizeof (struct mlx5_ib_create_qp ); -struct ib_qp_init_attr *attr = params->attr;struct ib_udata *udata = params->udata;size_t outlen = udata->outlen;size_t inlen = udata->inlen; params->outlen =min (outlen,sizeof (struct mlx5_ib_create_qp_resp )); -if (attr->qp_type ==IB_QPT_DRIVER ) { - params->inlen = (inlen < ucmd) ?0 : ucmd; - goto out; - } - + params->ucmd_size = ucmd;if (!params->is_rss_raw) { - params->inlen = ucmd; +/* User has old rdma-core, which doesn't support ECE */ +size_t min_inlen = + offsetof(struct mlx5_ib_create_qp , ece_options); + +/* +* We will check in check_ucmd_data() that user +* cleared everything after inlen. +*/ + params->inlen = (inlen < min_inlen) ?0 :min (inlen, ucmd); goto out; }
@@ -2733,13 +2742,14 @@ static int process_udata_size(struct mlx5_ib_dev *dev,return -EINVAL ; ucmd =sizeof (struct mlx5_ib_create_qp_rss ); + params->ucmd_size = ucmd;if (inlen > ucmd && !ib_is_udata_cleared (udata, ucmd, inlen - ucmd))return -EINVAL ; params->inlen =min (ucmd, inlen); out:if (!params->inlen) -mlx5_ib_dbg (dev,"udata is too small or not cleared\n" ); +mlx5_ib_dbg (dev,"udata is too small\n" );return (params->inlen) ?0 : -EINVAL ; }
@@ -2855,6 +2865,49 @@ static int mlx5_ib_destroy_dct(struct mlx5_ib_qp *mqp)return 0 ; } +static int check_ucmd_data (struct mlx5_ib_dev *dev, +struct mlx5_create_qp_params *params) +{ +struct ib_qp_init_attr *attr = params->attr; +struct ib_udata *udata = params->udata; +size_t size, last; +int ret; + +if (params->is_rss_raw) +/* +* These QPs don't have "reserved" field in their +* create_qp input struct, so their data is always valid. +*/ + last =sizeof (struct mlx5_ib_create_qp_rss ); +else +/* IB_QPT_RAW_PACKET and IB_QPT_DRIVER don't have ECE data */ +switch (attr->qp_type) { +case IB_QPT_DRIVER : +case IB_QPT_RAW_PACKET : + last = offsetof(struct mlx5_ib_create_qp , ece_options); +break ; +default : + last = offsetof(struct mlx5_ib_create_qp , reserved); + } + +if (udata->inlen <= last) +return 0 ; + +/* +* User provides different create_qp structures based on the +* flow and we need to know if he cleared memory after our +* struct create_qp ends. +*/ + size = udata->inlen - last; + ret =ib_is_udata_cleared (params->udata, last, size); +if (!ret) +mlx5_ib_dbg ( + dev, +"udata is not cleared, inlen = %lu, ucmd = %lu, last = %lu, size = %lu\n" , + udata->inlen, params->ucmd_size, last, size); +return ret ?0 : -EINVAL ; +} +struct ib_qp *mlx5_ib_create_qp (struct ib_pd *pd,struct ib_qp_init_attr *attr,struct ib_udata *udata) {
@@ -2888,7 +2941,11 @@ struct ib_qp *mlx5_ib_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attr,if (err)return ERR_PTR (err); - params.ucmd =kzalloc (params.inlen,GFP_KERNEL ); + err =check_ucmd_data (dev, ¶ms); +if (err) +return ERR_PTR (err); + + params.ucmd =kzalloc (params.ucmd_size,GFP_KERNEL );if (!params.ucmd)return ERR_PTR (-ENOMEM );
include/uapi/rdma/mlx5-abi.h +2 -0
@@ -322,6 +322,8 @@ struct mlx5_ib_create_qp {__aligned_u64 sq_buf_addr;__aligned_u64 access_key; }; +__u32 ece_options; +__u32 reserved; };/* RX Hash function flags */
Signature · 1 line
2.26.2
PATCH v2 6/9 RDMA/mlx5: Convert modify QP to use MLX5_SET macros
In reply to Leon RomanovskyFrom: Leon Romanovsky <[email protected]>
Instead of hand crafted mlx5_qp_context and mlx5_qp_path use common MLX5_SET() macros.
--- drivers/infiniband/hw/mlx5/qp.c | 203 +++++++++++++++----------------- include/linux/mlx5/qp.h | 66 ----------- 2 files changed, 97 insertions(+), 172 deletions(-)
drivers/infiniband/hw/mlx5/qp.c +97 -106
@@ -3028,14 +3028,13 @@ int mlx5_ib_destroy_qp(struct ib_qp *qp, struct ib_udata *udata)return 0 ; } -static int to_mlx5_access_flags (struct mlx5_ib_qp *qp, -const struct ib_qp_attr *attr, -int attr_mask,__be32 *hw_access_flags_be) +static int set_qpc_atomic_flags (struct mlx5_ib_qp *qp, +const struct ib_qp_attr *attr,int attr_mask, +void *qpc) { -u8 dest_rd_atomic; -u32 access_flags, hw_access_flags =0 ; -struct mlx5_ib_dev *dev =to_mdev (qp->ibqp.device); +u8 dest_rd_atomic; +u32 access_flags;if (attr_mask &IB_QP_MAX_DEST_RD_ATOMIC ) dest_rd_atomic = attr->max_dest_rd_atomic;
@@ -3050,8 +3049,8 @@ static int to_mlx5_access_flags(struct mlx5_ib_qp *qp,if (!dest_rd_atomic) access_flags &=IB_ACCESS_REMOTE_WRITE ; -if (access_flags &IB_ACCESS_REMOTE_READ ) - hw_access_flags |=MLX5_QP_BIT_RRE ; +MLX5_SET (qpc, qpc, rre, !!(access_flags &IB_ACCESS_REMOTE_READ )); +if (access_flags &IB_ACCESS_REMOTE_ATOMIC ) {int atomic_mode;
@@ -3059,15 +3058,11 @@ static int to_mlx5_access_flags(struct mlx5_ib_qp *qp,if (atomic_mode <0 )return -EOPNOTSUPP ; - hw_access_flags |=MLX5_QP_BIT_RAE ; - hw_access_flags |= atomic_mode <<MLX5_ATOMIC_MODE_OFFSET ; +MLX5_SET (qpc, qpc, rae,1 ); +MLX5_SET (qpc, qpc, atomic_mode, atomic_mode); } -if (access_flags &IB_ACCESS_REMOTE_WRITE ) - hw_access_flags |=MLX5_QP_BIT_RWE ; - - *hw_access_flags_be =cpu_to_be32 (hw_access_flags); - +MLX5_SET (qpc, qpc, rwe, !!(access_flags &IB_ACCESS_REMOTE_WRITE ));return 0 ; }
@@ -3147,26 +3142,22 @@ static int modify_raw_packet_tx_affinity(struct mlx5_core_dev *dev,return err; } -static void mlx5_set_path_udp_sport (struct mlx5_qp_path *path, -const struct rdma_ah_attr *ah, +static void mlx5_set_path_udp_sport (void *path,const struct rdma_ah_attr *ah,u32 lqpn,u32 rqpn) {u32 fl = ah->grh.flow_label; -u16 sport;if (!fl) fl =rdma_calc_flow_label (lqpn, rqpn); - sport =rdma_flow_label_to_udp_sport (fl); - path->udp_sport =cpu_to_be16 (sport); +MLX5_SET (ads, path, udp_sport,rdma_flow_label_to_udp_sport (fl)); }static int mlx5_set_path (struct mlx5_ib_dev *dev,struct mlx5_ib_qp *qp, -const struct rdma_ah_attr *ah, -struct mlx5_qp_path *path,u8 port,int attr_mask, -u32 path_flags,const struct ib_qp_attr *attr, -bool alt) +const struct rdma_ah_attr *ah,void *path,u8 port, +int attr_mask,u32 path_flags, +const struct ib_qp_attr *attr,bool alt) {const struct ib_global_route *grh =rdma_ah_read_grh (ah);int err;
@@ -3175,8 +3166,8 @@ static int mlx5_set_path(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,u8 sl =rdma_ah_get_sl (ah);if (attr_mask &IB_QP_PKEY_INDEX ) - path->pkey_index =cpu_to_be16 (alt ? attr->alt_pkey_index : - attr->pkey_index); +MLX5_SET (ads, path, pkey_index, + alt ? attr->alt_pkey_index : attr->pkey_index);if (ah_flags &IB_AH_GRH ) { if (grh->sgid_index >=
@@ -3192,7 +3183,8 @@ static int mlx5_set_path(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,if (!(ah_flags &IB_AH_GRH ))return -EINVAL ; -memcpy (path->rmac, ah->roce.dmac,sizeof (ah->roce.dmac)); +ether_addr_copy (MLX5_ADDR_OF (ads, path, rmac_47_32), + ah->roce.dmac); if ((qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC || qp->ibqp.qp_type == IB_QPT_XRC_INI ||
@@ -3202,38 +3194,38 @@ static int mlx5_set_path(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,mlx5_set_path_udp_sport (path, ah, qp->ibqp.qp_num, attr->dest_qp_num); - path->dci_cfi_prio_sl = (sl &0x7 ) <<4 ; +MLX5_SET (ads, path, eth_prio, sl &0x7 ); gid_type = ah->grh.sgid_attr->gid_type;if (gid_type ==IB_GID_TYPE_ROCE_UDP_ENCAP ) - path->ecn_dscp = (grh->traffic_class >>2 ) &0x3f ; +MLX5_SET (ads, path, dscp, grh->traffic_class >>2 ); }else { - path->fl_free_ar = (path_flags &MLX5_PATH_FLAG_FL ) ?0x80 :0 ; - path->fl_free_ar |= - (path_flags &MLX5_PATH_FLAG_FREE_AR ) ?0x40 :0 ; - path->rlid =cpu_to_be16 (rdma_ah_get_dlid (ah)); - path->grh_mlid =rdma_ah_get_path_bits (ah) &0x7f ; -if (ah_flags &IB_AH_GRH ) - path->grh_mlid |=1 <<7 ; - path->dci_cfi_prio_sl = sl &0xf ; +MLX5_SET (ads, path, fl, !!(path_flags &MLX5_PATH_FLAG_FL )); +MLX5_SET (ads, path, free_ar, + !!(path_flags &MLX5_PATH_FLAG_FREE_AR )); +MLX5_SET (ads, path, rlid,rdma_ah_get_dlid (ah)); +MLX5_SET (ads, path, mlid,rdma_ah_get_path_bits (ah)); +MLX5_SET (ads, path, grh, !!(ah_flags &IB_AH_GRH )); +MLX5_SET (ads, path, sl, sl); }if (ah_flags &IB_AH_GRH ) { - path->mgid_index = grh->sgid_index; - path->hop_limit = grh->hop_limit; - path->tclass_flowlabel = -cpu_to_be32 ((grh->traffic_class <<20 ) | - (grh->flow_label)); -memcpy (path->rgid, grh->dgid.raw,16 ); +MLX5_SET (ads, path, src_addr_index, grh->sgid_index); +MLX5_SET (ads, path, hop_limit, grh->hop_limit); +MLX5_SET (ads, path, tclass, grh->traffic_class); +MLX5_SET (ads, path, flow_label, grh->flow_label); +memcpy (MLX5_ADDR_OF (ads, path, rgid_rip), grh->dgid.raw, +sizeof (grh->dgid.raw)); } err =ib_rate_to_mlx5 (dev,rdma_ah_get_static_rate (ah));if (err <0 )return err; - path->static_rate = err; - path->port = port; +MLX5_SET (ads, path, stat_rate, err); +MLX5_SET (ads, path, vhca_port_num, port);if (attr_mask &IB_QP_TIMEOUT ) - path->ackto_lt = (alt ? attr->alt_timeout : attr->timeout) <<3 ; +MLX5_SET (ads, path, ack_timeout, + alt ? attr->alt_timeout : attr->timeout);if ((qp->ibqp.qp_type ==IB_QPT_RAW_PACKET ) && qp->sq.wqe_cnt) return modify_raw_packet_eth_prio(dev->mdev,
@@ -3759,9 +3751,9 @@ static int __mlx5_ib_modify_qp(struct ib_qp *ibqp,struct mlx5_ib_qp *qp =to_mqp (ibqp);struct mlx5_ib_qp_base *base = &qp->trans_qp.base;struct mlx5_ib_cq *send_cq, *recv_cq; -struct mlx5_qp_context *context;struct mlx5_ib_pd *pd;enum mlx5_qp_state mlx5_cur, mlx5_new; +void *qpc, *pri_path, *alt_path;enum mlx5_qp_optpar optpar =0 ;u32 set_id =0 ;int mlx5_st;
@@ -3773,25 +3765,25 @@ static int __mlx5_ib_modify_qp(struct ib_qp *ibqp,if (mlx5_st <0 )return -EINVAL ; - context =kzalloc (sizeof (*context),GFP_KERNEL ); -if (!context) + qpc =kzalloc (MLX5_ST_SZ_BYTES (qpc),GFP_KERNEL ); +if (!qpc)return -ENOMEM ; pd =to_mpd (qp->ibqp.pd); - context->flags =cpu_to_be32 (mlx5_st <<16 ); +MLX5_SET (qpc, qpc, st, mlx5_st);if (!(attr_mask &IB_QP_PATH_MIG_STATE )) { - context->flags |=cpu_to_be32 (MLX5_QP_PM_MIGRATED <<11 ); +MLX5_SET (qpc, qpc, pm_state,MLX5_QP_PM_MIGRATED ); }else {switch (attr->path_mig_state) {case IB_MIG_MIGRATED : - context->flags |=cpu_to_be32 (MLX5_QP_PM_MIGRATED <<11 ); +MLX5_SET (qpc, qpc, pm_state,MLX5_QP_PM_MIGRATED );break ;case IB_MIG_REARM : - context->flags |=cpu_to_be32 (MLX5_QP_PM_REARM <<11 ); +MLX5_SET (qpc, qpc, pm_state,MLX5_QP_PM_REARM );break ;case IB_MIG_ARMED : - context->flags |=cpu_to_be32 (MLX5_QP_PM_ARMED <<11 ); +MLX5_SET (qpc, qpc, pm_state,MLX5_QP_PM_ARMED );break ; } }
@@ -3799,19 +3791,20 @@ static int __mlx5_ib_modify_qp(struct ib_qp *ibqp, tx_affinity =get_tx_affinity (ibqp, attr, attr_mask, cur_state ==IB_QPS_RESET && new_state ==IB_QPS_INIT , udata); -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 ; - } + +MLX5_SET (qpc, qpc, lag_tx_port_affinity, tx_affinity); +if (tx_affinity && 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 ; +MLX5_SET (qpc, qpc, mtu,IB_MTU_256 ); +MLX5_SET (qpc, qpc, log_msg_max,8 ); }else if ((ibqp->qp_type ==IB_QPT_UD && !(qp->flags &IB_QP_CREATE_SOURCE_QPN )) || ibqp->qp_type ==MLX5_IB_QPT_REG_UMR ) { - context->mtu_msgmax = (IB_MTU_4096 <<5 ) |12 ; +MLX5_SET (qpc, qpc, mtu,IB_MTU_4096 ); +MLX5_SET (qpc, qpc, log_msg_max,12 ); }else if (attr_mask &IB_QP_PATH_MTU ) {if (attr->path_mtu <IB_MTU_256 || attr->path_mtu >IB_MTU_4096 ) {
@@ -3819,40 +3812,45 @@ static int __mlx5_ib_modify_qp(struct ib_qp *ibqp, err = -EINVAL ; goto out; } - context->mtu_msgmax = (attr->path_mtu <<5 ) | - (u8 )MLX5_CAP_GEN (dev->mdev, log_max_msg); +MLX5_SET (qpc, qpc, mtu, attr->path_mtu); +MLX5_SET (qpc, qpc, log_msg_max, +MLX5_CAP_GEN (dev->mdev, log_max_msg)); }if (attr_mask &IB_QP_DEST_QPN ) - context->log_pg_sz_remote_qpn =cpu_to_be32 (attr->dest_qp_num); +MLX5_SET (qpc, qpc, remote_qpn, attr->dest_qp_num); + + pri_path =MLX5_ADDR_OF (qpc, qpc, primary_address_path); + alt_path =MLX5_ADDR_OF (qpc, qpc, secondary_address_path);if (attr_mask &IB_QP_PKEY_INDEX ) - context->pri_path.pkey_index =cpu_to_be16 (attr->pkey_index); +MLX5_SET (ads, pri_path, pkey_index, attr->pkey_index);/* todo implement counter_index functionality */ if (is_sqp (ibqp->qp_type)) - context->pri_path.port = qp->port; +MLX5_SET (ads, pri_path, vhca_port_num, qp->port);if (attr_mask &IB_QP_PORT ) - context->pri_path.port = attr->port_num; +MLX5_SET (ads, pri_path, vhca_port_num, attr->port_num);if (attr_mask &IB_QP_AV ) { - err =mlx5_set_path (dev, qp, &attr->ah_attr, &context->pri_path, - attr_mask &IB_QP_PORT ? attr->port_num : qp->port, + err =mlx5_set_path (dev, qp, &attr->ah_attr, pri_path, + attr_mask &IB_QP_PORT ? attr->port_num : + qp->port, attr_mask,0 , attr, false);if (err) goto out; }if (attr_mask &IB_QP_TIMEOUT ) - context->pri_path.ackto_lt |= attr->timeout <<3 ; +MLX5_SET (ads, pri_path, ack_timeout, attr->timeout);if (attr_mask &IB_QP_ALT_PATH ) { - err =mlx5_set_path (dev, qp, &attr->alt_ah_attr, - &context->alt_path, + err =mlx5_set_path (dev, qp, &attr->alt_ah_attr, alt_path, attr->alt_port_num, - attr_mask |IB_QP_PKEY_INDEX |IB_QP_TIMEOUT , + attr_mask |IB_QP_PKEY_INDEX | +IB_QP_TIMEOUT ,0 , attr, true);if (err) goto out;
@@ -3861,53 +3859,47 @@ static int __mlx5_ib_modify_qp(struct ib_qp *ibqp,get_cqs (qp->ibqp.qp_type, qp->ibqp.send_cq, qp->ibqp.recv_cq, &send_cq, &recv_cq); - context->flags_pd =cpu_to_be32 (pd ? pd->pdn :to_mpd (dev->devr.p0)->pdn); - context->cqn_send = send_cq ?cpu_to_be32 (send_cq->mcq.cqn) :0 ; - context->cqn_recv = recv_cq ?cpu_to_be32 (recv_cq->mcq.cqn) :0 ; - context->params1 =cpu_to_be32 (MLX5_IB_ACK_REQ_FREQ <<28 ); +MLX5_SET (qpc, qpc, pd, pd ? pd->pdn :to_mpd (dev->devr.p0)->pdn); +if (send_cq) +MLX5_SET (qpc, qpc, cqn_snd, send_cq->mcq.cqn); +if (recv_cq) +MLX5_SET (qpc, qpc, cqn_rcv, recv_cq->mcq.cqn); + +MLX5_SET (qpc, qpc, log_ack_req_freq,MLX5_IB_ACK_REQ_FREQ );if (attr_mask &IB_QP_RNR_RETRY ) - context->params1 |=cpu_to_be32 (attr->rnr_retry <<13 ); +MLX5_SET (qpc, qpc, rnr_retry, attr->rnr_retry);if (attr_mask &IB_QP_RETRY_CNT ) - context->params1 |=cpu_to_be32 (attr->retry_cnt <<16 ); +MLX5_SET (qpc, qpc, retry_count, attr->retry_cnt); -if (attr_mask &IB_QP_MAX_QP_RD_ATOMIC ) { -if (attr->max_rd_atomic) - context->params1 |= -cpu_to_be32 (fls (attr->max_rd_atomic -1 ) <<21 ); - } +if (attr_mask &IB_QP_MAX_QP_RD_ATOMIC && attr->max_rd_atomic) +MLX5_SET (qpc, qpc, log_sra_max,ilog2 (attr->max_rd_atomic));if (attr_mask &IB_QP_SQ_PSN ) - context->next_send_psn =cpu_to_be32 (attr->sq_psn); +MLX5_SET (qpc, qpc, next_send_psn, attr->sq_psn); -if (attr_mask &IB_QP_MAX_DEST_RD_ATOMIC ) { -if (attr->max_dest_rd_atomic) - context->params2 |= -cpu_to_be32 (fls (attr->max_dest_rd_atomic -1 ) <<21 ); - } +if (attr_mask &IB_QP_MAX_DEST_RD_ATOMIC && attr->max_dest_rd_atomic) +MLX5_SET (qpc, qpc, log_rra_max, +ilog2 (attr->max_dest_rd_atomic));if (attr_mask & (IB_QP_ACCESS_FLAGS |IB_QP_MAX_DEST_RD_ATOMIC )) { -__be32 access_flags; - - err =to_mlx5_access_flags (qp, attr, attr_mask, &access_flags); + err =set_qpc_atomic_flags (qp, attr, attr_mask, qpc);if (err) goto out; - - context->params2 |= access_flags; }if (attr_mask &IB_QP_MIN_RNR_TIMER ) - context->rnr_nextrecvpsn |=cpu_to_be32 (attr->min_rnr_timer <<24 ); +MLX5_SET (qpc, qpc, min_rnr_nak, attr->min_rnr_timer);if (attr_mask &IB_QP_RQ_PSN ) - context->rnr_nextrecvpsn |=cpu_to_be32 (attr->rq_psn); +MLX5_SET (qpc, qpc, next_rcv_psn, attr->rq_psn);if (attr_mask &IB_QP_QKEY ) - context->qkey =cpu_to_be32 (attr->qkey); +MLX5_SET (qpc, qpc, q_key, attr->qkey);if (qp->rq.wqe_cnt && cur_state ==IB_QPS_RESET && new_state ==IB_QPS_INIT ) - context->db_rec_addr =cpu_to_be64 (qp->db.dma); +MLX5_SET64 (qpc, qpc, dbr_addr, qp->db.dma);if (cur_state ==IB_QPS_RESET && new_state ==IB_QPS_INIT ) { u8 port_num = (attr_mask & IB_QP_PORT ? attr->port_num :
@@ -3921,15 +3913,14 @@ static int __mlx5_ib_modify_qp(struct ib_qp *ibqp, set_id = ibqp->counter->id; else set_id =mlx5_ib_get_counters_id (dev, port_num); - context->qp_counter_set_usr_page |= -cpu_to_be32 (set_id <<24 ); +MLX5_SET (qpc, qpc, counter_set_id, set_id); }if (!ibqp->uobject && cur_state ==IB_QPS_RESET && new_state ==IB_QPS_INIT ) - context->sq_crq_size |=cpu_to_be16 (1 <<4 ); +MLX5_SET (qpc, qpc, rlky,1 );if (qp->flags &MLX5_IB_QP_CREATE_SQPN_QP1 ) - context->deth_sqpn =cpu_to_be32 (1 ); +MLX5_SET (qpc, qpc, deth_sqpn,1 ); mlx5_cur =to_mlx5_state (cur_state); mlx5_new =to_mlx5_state (new_state);
@@ -3987,7 +3978,7 @@ static int __mlx5_ib_modify_qp(struct ib_qp *ibqp, err =modify_raw_packet_qp (dev, qp, &raw_qp_param, tx_affinity); }else { - err =mlx5_core_qp_modify (dev, op, optpar, context, &base->mqp); + err =mlx5_core_qp_modify (dev, op, optpar, qpc, &base->mqp); }if (err)
@@ -4034,7 +4025,7 @@ static int __mlx5_ib_modify_qp(struct ib_qp *ibqp, } out: -kfree (context); +kfree (qpc);return err; }
include/linux/mlx5/qp.h +0 -66
72 lines not on this page · raw message
--
2.26.2
PATCH v2 8/9 RDMA/mlx5: Set ECE options during modify QP
In reply to Leon RomanovskyFrom: Leon Romanovsky <[email protected]>
The most common way to set ECE option will be during modify QP command in INIT2RTR, RTR2RTS and RTS2RTS stages, so update mlx5 to support it.
--- drivers/infiniband/hw/mlx5/qp.c | 21 ++++++++++++--------- drivers/infiniband/hw/mlx5/qp.h | 2 +- drivers/infiniband/hw/mlx5/qpc.c | 11 +++++++---- include/uapi/rdma/mlx5-abi.h | 2 +- 4 files changed, 21 insertions(+), 15 deletions(-)
drivers/infiniband/hw/mlx5/qp.c +12 -9
@@ -2353,7 +2353,7 @@ static void destroy_qp_common(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,if (qp->ibqp.qp_type !=IB_QPT_RAW_PACKET && !(qp->flags &IB_QP_CREATE_SOURCE_QPN )) { err =mlx5_core_qp_modify (dev,MLX5_CMD_OP_2RST_QP ,0 , -NULL , &base->mqp); +NULL , &base->mqp,NULL ); }else {struct mlx5_modify_raw_qp_param raw_qp_param = { .operation =MLX5_CMD_OP_2RST_QP
@@ -3978,7 +3978,10 @@ static int __mlx5_ib_modify_qp(struct ib_qp *ibqp, err =modify_raw_packet_qp (dev, qp, &raw_qp_param, tx_affinity); }else { - err =mlx5_core_qp_modify (dev, op, optpar, qpc, &base->mqp); +u32 ece =MLX5_CAP_GEN (dev->mdev, ece_support) ? + ucmd->ece_options :0 ; + err =mlx5_core_qp_modify (dev, op, optpar, qpc, &base->mqp, + &ece); }if (err)
@@ -4131,7 +4134,6 @@ static int mlx5_ib_modify_dct(struct ib_qp *ibqp, struct ib_qp_attr *attr, set_id =mlx5_ib_get_counters_id (dev, attr->port_num -1 );MLX5_SET (dctc, dctc, counter_set_id, set_id); - }else if (cur_state ==IB_QPS_INIT && new_state ==IB_QPS_RTR ) {struct mlx5_ib_modify_qp_resp resp = {};u32 out[MLX5_ST_SZ_DW (create_dct_out)] = {0 };
@@ -4182,7 +4184,6 @@ int mlx5_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,struct mlx5_ib_modify_qp ucmd = {};enum ib_qp_type qp_type;enum ib_qp_state cur_state, new_state; -size_t required_cmd_sz;int err = -EINVAL ;int port;
@@ -4190,9 +4191,7 @@ int mlx5_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,return -ENOSYS ;if (udata && udata->inlen) { - required_cmd_sz = offsetof(typeof (ucmd ), reserved) + -sizeof (ucmd.reserved); -if (udata->inlen < required_cmd_sz) +if (udata->inlen <offsetofend (typeof (ucmd), ece_options))return -EINVAL ; if (udata->inlen > sizeof(ucmd) &&
@@ -4205,10 +4204,10 @@ int mlx5_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,return -EFAULT ;if (ucmd.comp_mask || -memchr_inv (&ucmd.reserved,0 ,sizeof (ucmd.reserved)) ||memchr_inv (&ucmd.burst_info.reserved,0 ,sizeof (ucmd.burst_info.reserved)))return -EOPNOTSUPP ; + }if (unlikely (ibqp->qp_type ==IB_QPT_GSI ))
@@ -4217,8 +4216,12 @@ int mlx5_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, qp_type = (unlikely (ibqp->qp_type ==MLX5_IB_QPT_HW_GSI )) ?IB_QPT_GSI : qp->type; -if (qp_type ==MLX5_IB_QPT_DCT ) +if (qp_type ==MLX5_IB_QPT_DCT ) { +if (memchr_inv (&ucmd.ece_options,0 ,sizeof (ucmd.ece_options))) +return -EOPNOTSUPP ; +return mlx5_ib_modify_dct (ibqp, attr, attr_mask, udata); + }mutex_lock (&qp->mutex);
drivers/infiniband/hw/mlx5/qp.h +1 -1
@@ -16,7 +16,7 @@ int mlx5_core_create_dct(struct mlx5_ib_dev *dev, struct mlx5_core_dct *qp,int mlx5_qpc_create_qp (struct mlx5_ib_dev *dev,struct mlx5_core_qp *qp,u32 *in,int inlen,u32 *out);int mlx5_core_qp_modify (struct mlx5_ib_dev *dev,u16 opcode,u32 opt_param_mask, -void *qpc,struct mlx5_core_qp *qp); +void *qpc,struct mlx5_core_qp *qp,u32 *ece);int mlx5_core_destroy_qp (struct mlx5_ib_dev *dev,struct mlx5_core_qp *qp);int mlx5_core_destroy_dct (struct mlx5_ib_dev *dev,struct mlx5_core_dct *dct); int mlx5_core_qp_query(struct mlx5_ib_dev *dev, struct mlx5_core_qp *qp,
drivers/infiniband/hw/mlx5/qpc.c +7 -4
@@ -343,7 +343,7 @@ static void mbox_free(struct mbox_info *mbox)static int modify_qp_mbox_alloc (struct mlx5_core_dev *dev,u16 opcode,int qpn,u32 opt_param_mask,void *qpc, -struct mbox_info *mbox,u16 uid) +struct mbox_info *mbox,u16 uid,u32 ece) { mbox->out =NULL ; mbox->in =NULL ;
@@ -391,18 +391,21 @@ static int modify_qp_mbox_alloc(struct mlx5_core_dev *dev, u16 opcode, int qpn,return -ENOMEM ;MOD_QP_IN_SET_QPC (init2rtr_qp, mbox->in, opcode, qpn, opt_param_mask, qpc, uid); +MLX5_SET (init2rtr_qp_in, mbox->in, ece, ece);break ;case MLX5_CMD_OP_RTR2RTS_QP :if (MBOX_ALLOC (mbox, rtr2rts_qp))return -ENOMEM ;MOD_QP_IN_SET_QPC (rtr2rts_qp, mbox->in, opcode, qpn, opt_param_mask, qpc, uid); +MLX5_SET (rtr2rts_qp_in, mbox->in, ece, ece);break ;case MLX5_CMD_OP_RTS2RTS_QP :if (MBOX_ALLOC (mbox, rts2rts_qp))return -ENOMEM ;MOD_QP_IN_SET_QPC (rts2rts_qp, mbox->in, opcode, qpn, opt_param_mask, qpc, uid); +MLX5_SET (rts2rts_qp_in, mbox->in, ece, ece);break ;case MLX5_CMD_OP_SQERR2RTS_QP :if (MBOX_ALLOC (mbox, sqerr2rts_qp))
@@ -423,13 +426,13 @@ static int modify_qp_mbox_alloc(struct mlx5_core_dev *dev, u16 opcode, int qpn, }int mlx5_core_qp_modify (struct mlx5_ib_dev *dev,u16 opcode,u32 opt_param_mask, -void *qpc,struct mlx5_core_qp *qp) +void *qpc,struct mlx5_core_qp *qp,u32 *ece) {struct mbox_info mbox;int err; - err =modify_qp_mbox_alloc (dev->mdev, opcode, qp->qpn, - opt_param_mask, qpc, &mbox, qp->uid); + err =modify_qp_mbox_alloc (dev->mdev, opcode, qp->qpn, opt_param_mask, + qpc, &mbox, qp->uid, (ece) ? *ece :0 );if (err)return err;
include/uapi/rdma/mlx5-abi.h +1 -1
@@ -423,7 +423,7 @@ struct mlx5_ib_burst_info {struct mlx5_ib_modify_qp {__u32 comp_mask;struct mlx5_ib_burst_info burst_info; -__u32 reserved; +__u32 ece_options; };struct mlx5_ib_modify_qp_resp {
Signature · 1 line
2.26.2
PATCH v2 9/9 RDMA/mlx5: Return ECE data after modify QP
In reply to Leon RomanovskyFrom: Leon Romanovsky <[email protected]>
After users sets the ECE option, FW will return the agreed/supported bits through an output structures of modify QP stages for regular QPs or through create QP for the DCT.
--- drivers/infiniband/hw/mlx5/qp.c | 25 +++++++++++++++++++++---- drivers/infiniband/hw/mlx5/qpc.c | 25 +++++++++++++++++++++++++ include/uapi/rdma/mlx5-abi.h | 2 ++ 3 files changed, 48 insertions(+), 4 deletions(-)
drivers/infiniband/hw/mlx5/qp.c +21 -4
@@ -3708,6 +3708,7 @@ static int __mlx5_ib_modify_qp(struct ib_qp *ibqp, enum ib_qp_state cur_state, enum ib_qp_state new_state, const struct mlx5_ib_modify_qp *ucmd, + struct mlx5_ib_modify_qp_resp *resp, struct ib_udata *udata) { static const u16 optab[MLX5_QP_NUM_STATE][MLX5_QP_NUM_STATE] = {
@@ -3978,10 +3979,15 @@ static int __mlx5_ib_modify_qp(struct ib_qp *ibqp, err =modify_raw_packet_qp (dev, qp, &raw_qp_param, tx_affinity); }else { -u32 ece =MLX5_CAP_GEN (dev->mdev, ece_support) ? - ucmd->ece_options :0 ; +if (udata) { +/* For the kernel flows, the resp will stay zero */ + resp->ece_options = +MLX5_CAP_GEN (dev->mdev, ece_support) ? + ucmd->ece_options :0 ; + resp->response_length =sizeof (*resp); + } err =mlx5_core_qp_modify (dev, op, optpar, qpc, &base->mqp, - &ece); + &resp->ece_options); }if (err)
@@ -4180,6 +4186,7 @@ int mlx5_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask, struct ib_udata *udata) { struct mlx5_ib_dev *dev = to_mdev(ibqp->device); +struct mlx5_ib_modify_qp_resp resp = {};struct mlx5_ib_qp *qp =to_mqp (ibqp);struct mlx5_ib_modify_qp ucmd = {};enum ib_qp_type qp_type;
@@ -4292,7 +4299,17 @@ int mlx5_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, } err =__mlx5_ib_modify_qp (ibqp, attr, attr_mask, cur_state, - new_state, &ucmd, udata); + new_state, &ucmd, &resp, udata); + +/* resp.response_length is set in ECE supported flows only */ +if (!err && resp.response_length && + udata->outlen >= resp.response_length) +/* +* We don't check return value of the function below +* on purpose, because it is unclear how to unwind the +* error flow after QP was modified to the new state. +*/ +ib_copy_to_udata (udata, &resp, resp.response_length); out:mutex_unlock (&qp->mutex);
drivers/infiniband/hw/mlx5/qpc.c +25 -0
@@ -341,6 +341,27 @@ static void mbox_free(struct mbox_info *mbox)kfree (mbox->out); } +static int get_ece_from_mbox (void *out,u16 opcode) +{ +int ece =0 ; + +switch (opcode) { +case MLX5_CMD_OP_INIT2RTR_QP : + ece =MLX5_GET (init2rtr_qp_out, out, ece); +break ; +case MLX5_CMD_OP_RTR2RTS_QP : + ece =MLX5_GET (rtr2rts_qp_out, out, ece); +break ; +case MLX5_CMD_OP_RTS2RTS_QP : + ece =MLX5_GET (rts2rts_qp_out, out, ece); +break ; +default : +break ; + } + +return ece; +} +static int modify_qp_mbox_alloc (struct mlx5_core_dev *dev,u16 opcode,int qpn,u32 opt_param_mask,void *qpc,struct mbox_info *mbox,u16 uid,u32 ece)
@@ -438,6 +459,10 @@ int mlx5_core_qp_modify(struct mlx5_ib_dev *dev, u16 opcode, u32 opt_param_mask, err =mlx5_cmd_exec (dev->mdev, mbox.in, mbox.inlen, mbox.out, mbox.outlen); + +if (ece) + *ece =get_ece_from_mbox (mbox.out, opcode); +mbox_free (&mbox);return err; }
include/uapi/rdma/mlx5-abi.h +2 -0
@@ -429,6 +429,8 @@ struct mlx5_ib_modify_qp {struct mlx5_ib_modify_qp_resp {__u32 response_length;__u32 dctn; +__u32 ece_options; +__u32 reserved; };struct mlx5_ib_create_wq_resp {
Signature · 1 line
2.26.2
PATCH v2 7/9 RDMA/mlx5: Advertise ECE support
In reply to Leon RomanovskyFrom: Leon Romanovsky <[email protected]>
The ECE bits are configured through create_qp and modify_qp commands. While the create_qp() can be easily extended, it is not an easy task for the modify_qp().
The new bit in the comp_mask is needed to mark that kernel supports ECE and can receive data instead of "reserved" field in the struct mlx5_ib_modify_qp.
--- drivers/infiniband/hw/mlx5/main.c | 3 +++ include/uapi/rdma/mlx5-abi.h | 1 + 2 files changed, 4 insertions(+)
drivers/infiniband/hw/mlx5/main.c +3 -0
@@ -1971,6 +1971,9 @@ static int mlx5_ib_alloc_ucontext(struct ib_ucontext *uctx, resp.response_length +=sizeof (resp.dump_fill_mkey); } +if (MLX5_CAP_GEN (dev->mdev, ece_support)) + resp.comp_mask |=MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_ECE ; + err =ib_copy_to_udata (udata, &resp, resp.response_length);if (err) goto out_mdev;
include/uapi/rdma/mlx5-abi.h +1 -0
@@ -100,6 +100,7 @@ struct mlx5_ib_alloc_ucontext_req_v2 {enum mlx5_ib_alloc_ucontext_resp_mask {MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_CORE_CLOCK_OFFSET =1UL <<0 ,MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_DUMP_FILL_MKEY =1UL <<1 , +MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_ECE =1UL <<2 , };enum mlx5_user_cmds_supp_uhw {
--
2.26.2