From: Guy Levi <guyle@xxxxxxxxxxxx>
Add support to work with a RSS QP by using an indirection table
object and RX Hash attributes as a driver data upon QP creation.
Signed-off-by: Guy Levi <guyle@xxxxxxxxxxxx>
Reviewed-by: Yishai Hadas <yishaih@xxxxxxxxxxxx>
---
providers/mlx4/mlx4-abi.h | 14 ++++++
providers/mlx4/mlx4.h | 6 +++
providers/mlx4/verbs.c | 106 ++++++++++++++++++++++++++++++++++++++++++++++
3 files changed, 126 insertions(+)
diff --git a/providers/mlx4/mlx4-abi.h b/providers/mlx4/mlx4-abi.h
index 1cf6d5b..4f2132c 100644
--- a/providers/mlx4/mlx4-abi.h
+++ b/providers/mlx4/mlx4-abi.h
@@ -141,6 +141,20 @@ struct mlx4_create_qp {
__u32 inl_recv_sz;
};
+struct mlx4_create_qp_drv_ex_rss {
+ __u64 hash_fields_mask; /* enum ibv_rx_hash_fields */
+ __u8 hash_function; /* enum ibv_rx_hash_function_flags */
+ __u8 reserved[7];
+ __u8 hash_key[40];
+ __u32 comp_mask;
+ __u32 reserved1;
+};
+
+struct mlx4_create_qp_ex_rss {
+ struct ibv_create_qp_ex ibv_cmd;
+ struct mlx4_create_qp_drv_ex_rss drv_ex;
+};
+
struct mlx4_create_qp_drv_ex {
__u64 buf_addr;
__u64 db_addr;
diff --git a/providers/mlx4/mlx4.h b/providers/mlx4/mlx4.h
index 13a561f..83dd551 100644
--- a/providers/mlx4/mlx4.h
+++ b/providers/mlx4/mlx4.h
@@ -198,6 +198,11 @@ struct mlx4_wq {
int offset;
};
+enum mlx4_rsc_type {
+ MLX4_RSC_TYPE_QP = 0,
+ MLX4_RSC_TYPE_RSS_QP = 1,
+};
+
struct mlx4_qp {
union {
struct verbs_qp verbs_qp;
@@ -216,6 +221,7 @@ struct mlx4_qp {
struct mlx4_wq rq;
uint8_t link_layer;
+ uint8_t type; /* enum mlx4_rsc_type */
uint32_t qp_cap_cache;
};
diff --git a/providers/mlx4/verbs.c b/providers/mlx4/verbs.c
index 369f437..5f2266a 100644
--- a/providers/mlx4/verbs.c
+++ b/providers/mlx4/verbs.c
@@ -745,6 +745,70 @@ int mlx4_destroy_srq(struct ibv_srq *srq)
return 0;
}
+static int mlx4_cmd_create_qp_ex_rss(struct ibv_context *context,
+ struct ibv_qp_init_attr_ex *attr,
+ struct mlx4_create_qp *cmd,
+ struct mlx4_qp *qp)
+{
+ struct mlx4_create_qp_ex_rss cmd_ex = {};
+ struct mlx4_create_qp_resp_ex resp;
+ int ret;
+
+ if (attr->rx_hash_conf.rx_hash_key_len !=
+ sizeof(cmd_ex.drv_ex.hash_key)) {
+ errno = ENOTSUP;
+ return errno;
+ }
+
+ cmd_ex.drv_ex.hash_fields_mask =
+ attr->rx_hash_conf.rx_hash_fields_mask;
+ cmd_ex.drv_ex.hash_function =
+ attr->rx_hash_conf.rx_hash_function;
+ memcpy(cmd_ex.drv_ex.hash_key, attr->rx_hash_conf.rx_hash_key,
+ sizeof(cmd_ex.drv_ex.hash_key));
+
+ ret = ibv_cmd_create_qp_ex2(context, &qp->verbs_qp,
+ sizeof(qp->verbs_qp), attr,
+ &cmd_ex.ibv_cmd, sizeof(cmd_ex.ibv_cmd),
+ sizeof(cmd_ex), &resp.ibv_resp,
+ sizeof(resp.ibv_resp), sizeof(resp));
+ return ret;
+}
+
+static struct ibv_qp *_mlx4_create_qp_ex_rss(struct ibv_context *context,
+ struct ibv_qp_init_attr_ex *attr)
+{
+ struct mlx4_create_qp cmd = {};
+ struct mlx4_qp *qp;
+ int ret;
+
+ if (!(attr->comp_mask & IBV_QP_INIT_ATTR_RX_HASH) ||
+ !(attr->comp_mask & IBV_QP_INIT_ATTR_IND_TABLE))
+ return NULL;
+
+ if (attr->qp_type != IBV_QPT_RAW_PACKET)
+ return NULL;
+
+ qp = calloc(1, sizeof(*qp));
+ if (!qp)
+ return NULL;
+
+ if (pthread_spin_init(&qp->sq.lock, PTHREAD_PROCESS_PRIVATE) ||
+ pthread_spin_init(&qp->rq.lock, PTHREAD_PROCESS_PRIVATE))
+ goto err;
+
+ ret = mlx4_cmd_create_qp_ex_rss(context, attr, &cmd, qp);
+ if (ret)
+ goto err;
+
+ qp->type = MLX4_RSC_TYPE_RSS_QP;
+
+ return &qp->verbs_qp.qp;
+err:
+ free(qp);
+ return NULL;
+}
+
static int mlx4_cmd_create_qp_ex(struct ibv_context *context,
struct ibv_qp_init_attr_ex *attr,
struct mlx4_create_qp *cmd,
@@ -792,6 +856,11 @@ static struct ibv_qp *create_qp_ex(struct ibv_context *context,
struct mlx4_qp *qp;
int ret;
+ if (attr->comp_mask & (IBV_QP_INIT_ATTR_RX_HASH |
+ IBV_QP_INIT_ATTR_IND_TABLE)) {
+ return _mlx4_create_qp_ex_rss(context, attr);
+ }
+
/* Sanity check QP size before proceeding */
if (ctx->max_qp_wr) { /* mlx4_query_device succeeded */
if (attr->cap.max_send_wr > ctx->max_qp_wr ||
@@ -987,6 +1056,9 @@ int mlx4_query_qp(struct ibv_qp *ibqp, struct ibv_qp_attr *attr,
struct mlx4_qp *qp = to_mqp(ibqp);
int ret;
+ if (qp->type == MLX4_RSC_TYPE_RSS_QP)
+ return ENOTSUP;
+
ret = ibv_cmd_query_qp(ibqp, attr, attr_mask, init_attr, &cmd, sizeof cmd);
if (ret)
return ret;
@@ -1000,6 +1072,20 @@ int mlx4_query_qp(struct ibv_qp *ibqp, struct ibv_qp_attr *attr,
return 0;
}
+static int _mlx4_modify_qp_rss(struct ibv_qp *qp, struct ibv_qp_attr *attr,
+ int attr_mask)
+{
+ struct ibv_modify_qp cmd = {};
+
+ if (attr_mask & ~(IBV_QP_STATE | IBV_QP_PORT))
+ return ENOTSUP;
+
+ if (attr->qp_state > IBV_QPS_RTR)
+ return ENOTSUP;
+
+ return ibv_cmd_modify_qp(qp, attr, attr_mask, &cmd, sizeof(cmd));
+}
+
int mlx4_modify_qp(struct ibv_qp *qp, struct ibv_qp_attr *attr,
int attr_mask)
{
@@ -1009,6 +1095,9 @@ int mlx4_modify_qp(struct ibv_qp *qp, struct ibv_qp_attr *attr,
struct ibv_device_attr device_attr;
int ret;
+ if (mqp->type == MLX4_RSC_TYPE_RSS_QP)
+ return _mlx4_modify_qp_rss(qp, attr, attr_mask);
+
memset(&device_attr, 0, sizeof(device_attr));
if (attr_mask & IBV_QP_PORT) {
ret = ibv_query_port(qp->context, attr->port_num,
@@ -1108,11 +1197,28 @@ static void mlx4_unlock_cqs(struct ibv_qp *qp)
}
}
+static int _mlx4_destroy_qp_rss(struct ibv_qp *ibqp)
+{
+ struct mlx4_qp *qp = to_mqp(ibqp);
+ int ret;
+
+ ret = ibv_cmd_destroy_qp(ibqp);
+ if (ret && !cleanup_on_fatal(ret))
+ return ret;
+
+ free(qp);
+
+ return 0;
+}
+
int mlx4_destroy_qp(struct ibv_qp *ibqp)
{
struct mlx4_qp *qp = to_mqp(ibqp);
int ret;
+ if (qp->type == MLX4_RSC_TYPE_RSS_QP)
+ return _mlx4_destroy_qp_rss(ibqp);
+
pthread_mutex_lock(&to_mctx(ibqp->context)->qp_table_mutex);
ret = ibv_cmd_destroy_qp(ibqp);
if (ret && !cleanup_on_fatal(ret)) {
--
1.8.3.1
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
