RX checksum verification status is reported through wc_flag when polling CQ if
device supports checksum offload. When IBV_WC_IP_CSUM_OK is set, that means
both IPv4 header checksum and TCP/UDP checksum are OK.
TX checksum offload will be enabled for TCP/UDP over IPv4 if user sets
send_flag IBV_SEND_IP_CSUM and device supports checksum offload.
A new field: qp_cap_cache, is added to mlx4_qp in order to 'cache' the device
capabilities to minimize performance hit on poll_one and post_send function.
The capabilities are set inside mlx4_modify_qp. Post_send will return error
if device doesn't support checksum but user sets flag IBV_SEND_IP_CSUM.
Signed-off-by: Bodong Wang <bod...@mellanox.com>
---
src/cq.c | 6 ++++++
src/mlx4.c | 1 +
src/mlx4.h | 23 ++++++++++++++++++++++-
src/qp.c | 19 +++++++++++++++++++
src/verbs.c | 54 ++++++++++++++++++++++++++++++++++++++++++++++++++++++
src/wqe.h | 8 +++++---
6 files changed, 107 insertions(+), 4 deletions(-)
diff --git a/src/cq.c b/src/cq.c
index 8b27795..32c9070 100644
--- a/src/cq.c
+++ b/src/cq.c
@@ -329,6 +329,12 @@ static int mlx4_poll_one(struct mlx4_cq *cq,
wc->sl = ntohs(cqe->sl_vid) >> 13;
else
wc->sl = ntohs(cqe->sl_vid) >> 12;
+
+ if ((*cur_qp) && ((*cur_qp)->qp_cap_cache &
MLX4_RX_CSUM_VALID)) {
+ wc->wc_flags |= ((cqe->status &
htonl(MLX4_CQE_STATUS_IPV4_CSUM_OK)) ==
+ htonl(MLX4_CQE_STATUS_IPV4_CSUM_OK)) <<
+ IBV_WC_IP_CSUM_OK_SHIFT;
+ }
}
return CQ_OK;
diff --git a/src/mlx4.c b/src/mlx4.c
index 9fe8c6a..427a3a8 100644
--- a/src/mlx4.c
+++ b/src/mlx4.c
@@ -205,6 +205,7 @@ static int mlx4_init_context(struct verbs_device *v_device,
verbs_set_ctx_op(verbs_ctx, open_qp, mlx4_open_qp);
verbs_set_ctx_op(verbs_ctx, ibv_create_flow, ibv_cmd_create_flow);
verbs_set_ctx_op(verbs_ctx, ibv_destroy_flow, ibv_cmd_destroy_flow);
+ verbs_set_ctx_op(verbs_ctx, query_device_ex, mlx4_query_device_ex);
return 0;
diff --git a/src/mlx4.h b/src/mlx4.h
index d71450f..7e229d7 100644
--- a/src/mlx4.h
+++ b/src/mlx4.h
@@ -257,6 +257,7 @@ struct mlx4_qp {
struct mlx4_wq rq;
uint8_t link_layer;
+ uint32_t qp_cap_cache;
};
struct mlx4_av {
@@ -279,6 +280,22 @@ struct mlx4_ah {
uint8_t mac[6];
};
+enum {
+ MLX4_CSUM_SUPPORT_UD_OVER_IB = (1 << 0),
+ MLX4_CSUM_SUPPORT_RAW_OVER_ETH = (1 << 1),
+ /* Only report rx checksum when the validation is valid */
+ MLX4_RX_CSUM_VALID = (1 << 16),
+};
+
+enum mlx4_cqe_status {
+ MLX4_CQE_STATUS_TCP_UDP_CSUM_OK = (1 << 2),
+ MLX4_CQE_STATUS_IPV4_PKT = (1 << 22),
+ MLX4_CQE_STATUS_IP_HDR_CSUM_OK = (1 << 28),
+ MLX4_CQE_STATUS_IPV4_CSUM_OK = MLX4_CQE_STATUS_IPV4_PKT |
+ MLX4_CQE_STATUS_IP_HDR_CSUM_OK |
+ MLX4_CQE_STATUS_TCP_UDP_CSUM_OK
+};
+
struct mlx4_cqe {
uint32_t vlan_my_qpn;
uint32_t immed_rss_invalid;
@@ -286,7 +303,7 @@ struct mlx4_cqe {
uint8_t sl_vid;
uint8_t reserved1;
uint16_t rlid;
- uint32_t reserved2;
+ uint32_t status;
uint32_t byte_cnt;
uint16_t wqe_index;
uint16_t checksum;
@@ -352,6 +369,10 @@ void mlx4_free_db(struct mlx4_context *context, enum
mlx4_db_type type, uint32_t
int mlx4_query_device(struct ibv_context *context,
struct ibv_device_attr *attr);
+int mlx4_query_device_ex(struct ibv_context *context,
+ const struct ibv_query_device_ex_input *input,
+ struct ibv_device_attr_ex *attr,
+ size_t attr_size);
int mlx4_query_port(struct ibv_context *context, uint8_t port,
struct ibv_port_attr *attr);
diff --git a/src/qp.c b/src/qp.c
index 721bed4..057490b 100644
--- a/src/qp.c
+++ b/src/qp.c
@@ -289,12 +289,31 @@ int mlx4_post_send(struct ibv_qp *ibqp, struct
ibv_send_wr *wr,
set_datagram_seg(wqe, wr);
wqe += sizeof (struct mlx4_wqe_datagram_seg);
size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
+
+ if (wr->send_flags & IBV_SEND_IP_CSUM) {
+ if (!(qp->qp_cap_cache &
MLX4_CSUM_SUPPORT_UD_OVER_IB)) {
+ ret = EINVAL;
+ *bad_wr = wr;
+ goto out;
+ }
+ ctrl->srcrb_flags |=
htonl(MLX4_WQE_CTRL_IP_HDR_CSUM |
+
MLX4_WQE_CTRL_TCP_UDP_CSUM);
+ }
break;
case IBV_QPT_RAW_PACKET:
/* For raw eth, the MLX4_WQE_CTRL_SOLICIT flag is used
* to indicate that no icrc should be calculated */
ctrl->srcrb_flags |= htonl(MLX4_WQE_CTRL_SOLICIT);
+ if (wr->send_flags & IBV_SEND_IP_CSUM) {
+ if (!(qp->qp_cap_cache &
MLX4_CSUM_SUPPORT_RAW_OVER_ETH)) {
+ ret = EINVAL;
+ *bad_wr = wr;
+ goto out;
+ }
+ ctrl->srcrb_flags |=
htonl(MLX4_WQE_CTRL_IP_HDR_CSUM |
+
MLX4_WQE_CTRL_TCP_UDP_CSUM);
+ }
break;
default:
diff --git a/src/verbs.c b/src/verbs.c
index 623d576..18ee786 100644
--- a/src/verbs.c
+++ b/src/verbs.c
@@ -66,6 +66,34 @@ int mlx4_query_device(struct ibv_context *context, struct
ibv_device_attr *attr)
return 0;
}
+int mlx4_query_device_ex(struct ibv_context *context,
+ const struct ibv_query_device_ex_input *input,
+ struct ibv_device_attr_ex *attr,
+ size_t attr_size)
+{
+ struct ibv_query_device_ex cmd;
+ struct ibv_query_device_resp_ex resp;
+ uint64_t raw_fw_ver;
+ unsigned major, minor, sub_minor;
+ int ret;
+
+ ret = ibv_cmd_query_device_ex(context, input, attr, attr_size,
+ &raw_fw_ver,
+ &cmd, sizeof(cmd), sizeof(cmd),
+ &resp, sizeof(resp), sizeof(resp));
+ if (ret)
+ return ret;
+
+ major = (raw_fw_ver >> 32) & 0xffff;
+ minor = (raw_fw_ver >> 16) & 0xffff;
+ sub_minor = raw_fw_ver & 0xffff;
+
+ snprintf(attr->orig_attr.fw_ver, sizeof attr->orig_attr.fw_ver,
+ "%d.%d.%03d", major, minor, sub_minor);
+
+ return 0;
+}
+
int mlx4_query_port(struct ibv_context *context, uint8_t port,
struct ibv_port_attr *attr)
{
@@ -606,14 +634,40 @@ int mlx4_modify_qp(struct ibv_qp *qp, struct ibv_qp_attr
*attr,
struct ibv_modify_qp cmd;
struct ibv_port_attr port_attr;
struct mlx4_qp *mqp = to_mqp(qp);
+ struct ibv_device_attr_ex device_attr;
+ struct ibv_query_device_ex_input input;
int ret;
+ memset(&device_attr, 0, sizeof(device_attr));
+ memset(&input, 0, sizeof(input));
if (attr_mask & IBV_QP_PORT) {
ret = ibv_query_port(qp->context, attr->port_num,
&port_attr);
if (ret)
return ret;
mqp->link_layer = port_attr.link_layer;
+
+ ret = ibv_query_device_ex(qp->context, &input, &device_attr);
+ if (ret)
+ return ret;
+
+ switch(qp->qp_type) {
+ case IBV_QPT_UD:
+ if ((mqp->link_layer == IBV_LINK_LAYER_INFINIBAND) &&
+ (device_attr.csum_cap.ib_csum_cap &
IBV_CSUM_SUPPORT_UD))
+ mqp->qp_cap_cache |=
MLX4_CSUM_SUPPORT_UD_OVER_IB |
+ MLX4_RX_CSUM_VALID;
+ break;
+ case IBV_QPT_RAW_PACKET:
+ if ((mqp->link_layer == IBV_LINK_LAYER_ETHERNET) &&
+ (device_attr.csum_cap.eth_csum_cap &
IBV_CSUM_SUPPORT_RAW))
+ mqp->qp_cap_cache |=
MLX4_CSUM_SUPPORT_RAW_OVER_ETH |
+ MLX4_RX_CSUM_VALID;
+ break;
+ default:
+ break;
+ }
+
}
if (qp->state == IBV_QPS_RESET &&
diff --git a/src/wqe.h b/src/wqe.h
index bbd22ba..bbfd7df 100644
--- a/src/wqe.h
+++ b/src/wqe.h
@@ -38,9 +38,11 @@ enum {
};
enum {
- MLX4_WQE_CTRL_FENCE = 1 << 6,
- MLX4_WQE_CTRL_CQ_UPDATE = 3 << 2,
- MLX4_WQE_CTRL_SOLICIT = 1 << 1,
+ MLX4_WQE_CTRL_FENCE = 1 << 6,
+ MLX4_WQE_CTRL_CQ_UPDATE = 3 << 2,
+ MLX4_WQE_CTRL_SOLICIT = 1 << 1,
+ MLX4_WQE_CTRL_IP_HDR_CSUM = 1 << 4,
+ MLX4_WQE_CTRL_TCP_UDP_CSUM = 1 << 5,
};
enum {
--
1.8.3.1
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
