On 2018/7/20 18:26, Yuval Shaia wrote:
> On Thu, Jul 19, 2018 at 08:23:17PM +0800, Yixian Liu wrote:
>> The cqe should be flushed error completion status if an related
>> error is detected while poll cqe, post send or post recv.
>>
>> Record doorbell is used to notify the head pointer of sq and rq
>> to the kernel.
>>
>> Signed-off-by: Yixian Liu <liuyixian@xxxxxxxxxx>
>> ---
>> kernel-headers/rdma/hns-abi.h | 1 +
>> providers/hns/hns_roce_u.h | 1 +
>> providers/hns/hns_roce_u_hw_v2.c | 53 ++++++++++++++++++++++++++++++++++++++++
>> providers/hns/hns_roce_u_hw_v2.h | 1 +
>> providers/hns/hns_roce_u_verbs.c | 24 +++++++++++++++---
>> 5 files changed, 77 insertions(+), 3 deletions(-)
>>
>> diff --git a/kernel-headers/rdma/hns-abi.h b/kernel-headers/rdma/hns-abi.h
>> index 78613b6..c1f8773 100644
>> --- a/kernel-headers/rdma/hns-abi.h
>> +++ b/kernel-headers/rdma/hns-abi.h
>> @@ -53,6 +53,7 @@ struct hns_roce_ib_create_qp {
>> __u8 log_sq_stride;
>> __u8 sq_no_prefetch;
>> __u8 reserved[5];
>> + __aligned_u64 sdb_addr;
>> };
>>
>> struct hns_roce_ib_create_qp_resp {
>> diff --git a/providers/hns/hns_roce_u.h b/providers/hns/hns_roce_u.h
>> index 8426569..2b2070f 100644
>> --- a/providers/hns/hns_roce_u.h
>> +++ b/providers/hns/hns_roce_u.h
>> @@ -211,6 +211,7 @@ struct hns_roce_qp {
>> struct hns_roce_wq sq;
>> struct hns_roce_wq rq;
>> uint32_t *rdb;
>> + uint32_t *sdb;
>> struct hns_roce_sge_ex sge;
>> unsigned int next_sge;
>> int port_num;
>> diff --git a/providers/hns/hns_roce_u_hw_v2.c b/providers/hns/hns_roce_u_hw_v2.c
>> index ca59011..588a946 100644
>> --- a/providers/hns/hns_roce_u_hw_v2.c
>> +++ b/providers/hns/hns_roce_u_hw_v2.c
>> @@ -237,6 +237,9 @@ static void hns_roce_v2_clear_qp(struct hns_roce_context *ctx, uint32_t qpn)
>> ctx->qp_table[tind].table[qpn & ctx->qp_table_mask] = NULL;
>> }
>>
>> +static int hns_roce_u_v2_modify_qp(struct ibv_qp *qp, struct ibv_qp_attr *attr,
>> + int attr_mask);
>> +
>> static int hns_roce_v2_poll_one(struct hns_roce_cq *cq,
>> struct hns_roce_qp **cur_qp, struct ibv_wc *wc)
>> {
>> @@ -248,6 +251,9 @@ static int hns_roce_v2_poll_one(struct hns_roce_cq *cq,
>> struct hns_roce_v2_cqe *cqe = NULL;
>> struct hns_roce_rinl_sge *sge_list;
>> uint32_t opcode;
>> + struct ibv_qp_attr attr;
>> + int attr_mask;
>> + int ret;
>>
>> /* According to CI, find the relative cqe */
>> cqe = next_cqe_sw_v2(cq);
>> @@ -314,6 +320,19 @@ static int hns_roce_v2_poll_one(struct hns_roce_cq *cq,
>> if (roce_get_field(cqe->byte_4, CQE_BYTE_4_STATUS_M,
>> CQE_BYTE_4_STATUS_S) != HNS_ROCE_V2_CQE_SUCCESS) {
>> hns_roce_v2_handle_error_cqe(cqe, wc);
>> +
>> + /* flush cqe */
>> + if ((wc->status != IBV_WC_SUCCESS) &&
>> + (wc->status != IBV_WC_WR_FLUSH_ERR)) {
>> + attr_mask = IBV_QP_STATE;
>> + attr.qp_state = IBV_QPS_ERR;
>> + ret = hns_roce_u_v2_modify_qp(&(*cur_qp)->ibv_qp,
>> + &attr, attr_mask);
>> + if (ret) {
>> + fprintf(stderr, PFX "failed to modify qp!\n");
>
> I do not understand why poll_one and got the honor of printing while
> post_send and post_recv are not.
>
> Is it because of caller taking care of it?
>
> Suggesting consistency.
>
Hi Yuval,
Thanks for your comment! I will fix it and keep consistent in next version.
--
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
