From: Matan Barak <[email protected]> This patch adds the new extended support for query_port. The purpose of this is: 1. Request fields that aren't availible by today's ibv_query_port 2. Don't fetch fields that the user doesn't need. Hence, there is more chance to optimize. 3. Cache link layer's type in mlx4_context. Caching will allow us to avoid ibv_query_port calls and save time in ibv_create_ah.
Signed-off-by: Matan Barak <[email protected]> Signed-off-by: Or Gerlitz <[email protected]> --- src/mlx4.c | 3 ++ src/mlx4.h | 9 ++++++++ src/verbs.c | 66 +++++++++++++++++++++++++++++++++++++++++++++++++++++----- 3 files changed, 72 insertions(+), 6 deletions(-) diff --git a/src/mlx4.c b/src/mlx4.c index a9f3a98..1d42270 100644 --- a/src/mlx4.c +++ b/src/mlx4.c @@ -157,6 +157,8 @@ static int mlx4_init_context(struct verbs_device *v_device, context->qp_table_shift = ffs(context->num_qps) - 1 - MLX4_QP_TABLE_BITS; context->qp_table_mask = (1 << context->qp_table_shift) - 1; + for (i = 0; i < MLX4_PORTS_NUM; ++i) + context->port_query_cache[i].valid = 0; pthread_mutex_init(&context->qp_table_mutex, NULL); for (i = 0; i < MLX4_QP_TABLE_SIZE; ++i) @@ -206,6 +208,7 @@ static int mlx4_init_context(struct verbs_device *v_device, verbs_set_ctx_op(verbs_ctx, create_flow, ibv_cmd_create_flow); verbs_set_ctx_op(verbs_ctx, destroy_flow, ibv_cmd_destroy_flow); verbs_set_ctx_op(verbs_ctx, create_ah_ex, mlx4_create_ah_ex); + verbs_set_ctx_op(verbs_ctx, query_port_ex, mlx4_query_port_ex); return 0; diff --git a/src/mlx4.h b/src/mlx4.h index 3015357..06fd2ba 100644 --- a/src/mlx4.h +++ b/src/mlx4.h @@ -40,6 +40,8 @@ #include <infiniband/arch.h> #include <infiniband/verbs.h> +#define MLX4_PORTS_NUM 2 + #ifdef HAVE_VALGRIND_MEMCHECK_H # include <valgrind/memcheck.h> @@ -189,6 +191,11 @@ struct mlx4_context { pthread_mutex_t db_list_mutex; int cqe_size; struct mlx4_xsrq_table xsrq_table; + struct { + uint8_t valid; + uint8_t link_layer; + enum ibv_port_cap_flags caps; + } port_query_cache[MLX4_PORTS_NUM]; }; struct mlx4_buf { @@ -354,6 +361,8 @@ int mlx4_query_device(struct ibv_context *context, struct ibv_device_attr *attr); int mlx4_query_port(struct ibv_context *context, uint8_t port, struct ibv_port_attr *attr); +int mlx4_query_port_ex(struct ibv_context *context, uint8_t port_num, + struct ibv_port_attr_ex *port_attr); struct ibv_pd *mlx4_alloc_pd(struct ibv_context *context); int mlx4_free_pd(struct ibv_pd *pd); diff --git a/src/verbs.c b/src/verbs.c index ccfd678..b2c71e7 100644 --- a/src/verbs.c +++ b/src/verbs.c @@ -72,8 +72,58 @@ int mlx4_query_port(struct ibv_context *context, uint8_t port, struct ibv_port_attr *attr) { struct ibv_query_port cmd; + int err; + + err = ibv_cmd_query_port(context, port, attr, &cmd, sizeof(cmd)); + if (!err && port <= MLX4_PORTS_NUM && port > 0) { + struct mlx4_context *mctx = to_mctx(context); + if (!mctx->port_query_cache[port - 1].valid) { + mctx->port_query_cache[port - 1].link_layer = + attr->link_layer; + mctx->port_query_cache[port - 1].caps = + attr->port_cap_flags; + mctx->port_query_cache[port - 1].valid = 1; + } + } + + return err; +} + +int mlx4_query_port_ex(struct ibv_context *context, uint8_t port_num, + struct ibv_port_attr_ex *port_attr) +{ + /* Check that only valid flags were given */ + if (port_attr->comp_mask & ~IBV_QUERY_PORT_EX_MASK) + return -EINVAL; + + /* Optimize the link type query */ + if (!(port_attr->comp_mask & ~(IBV_QUERY_PORT_EX_LINK_LAYER | + IBV_QUERY_PORT_EX_CAP_FLAGS))) { + struct mlx4_context *mctx = to_mctx(context); + if (port_num <= 0 || port_num > MLX4_PORTS_NUM) + return -EINVAL; + if (mctx->port_query_cache[port_num - 1].valid) { + if (port_attr->comp_mask & + IBV_QUERY_PORT_EX_LINK_LAYER) + port_attr->link_layer = + mctx-> + port_query_cache[port_num - 1]. + link_layer; + if (port_attr->comp_mask & + IBV_QUERY_PORT_EX_CAP_FLAGS) + port_attr->port_cap_flags = + mctx-> + port_query_cache[port_num - 1]. + caps; + return 0; + } + } + if (port_attr->comp_mask & IBV_QUERY_PORT_EX_STD_MASK) { + return mlx4_query_port(context, port_num, + (struct ibv_port_attr *)port_attr); + } - return ibv_cmd_query_port(context, port, attr, &cmd, sizeof cmd); + return -EOPNOTSUPP; } struct ibv_pd *mlx4_alloc_pd(struct ibv_context *context) @@ -826,15 +876,17 @@ static struct ibv_ah *mlx4_create_ah_common(struct ibv_pd *pd, struct ibv_ah *mlx4_create_ah(struct ibv_pd *pd, struct ibv_ah_attr *attr) { struct ibv_ah *ah; - struct ibv_port_attr port_attr; + struct ibv_port_attr_ex port_attr; + + port_attr.comp_mask = IBV_QUERY_PORT_EX_LINK_LAYER; - if (ibv_query_port(pd->context, attr->port_num, &port_attr)) + if (ibv_query_port_ex(pd->context, attr->port_num, &port_attr)) return NULL; ah = mlx4_create_ah_common(pd, attr, port_attr.link_layer); if (NULL != ah && (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET || - !mlx4_resolve_grh_to_l2(pd, to_mah(ah), attr))) + !mlx4_resolve_grh_to_l2(pd, to_mah(ah), attr))) return ah; if (ah) @@ -845,11 +897,13 @@ struct ibv_ah *mlx4_create_ah(struct ibv_pd *pd, struct ibv_ah_attr *attr) struct ibv_ah *mlx4_create_ah_ex(struct ibv_pd *pd, struct ibv_ah_attr_ex *attr_ex) { - struct ibv_port_attr port_attr; + struct ibv_port_attr_ex port_attr; struct ibv_ah *ah; struct mlx4_ah *mah; - if (ibv_query_port(pd->context, attr_ex->port_num, &port_attr)) + port_attr.comp_mask = IBV_QUERY_PORT_EX_LINK_LAYER; + + if (ibv_query_port_ex(pd->context, attr_ex->port_num, &port_attr)) return NULL; ah = mlx4_create_ah_common(pd, (struct ibv_ah_attr *)attr_ex, -- 1.7.1 -- To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html
