On 27/05/2020 15:53, Jason Gunthorpe wrote:
> On Wed, May 27, 2020 at 03:34:33PM +0300, Gal Pressman wrote:
>> Hi,
>>
>> The recent transition of create/destroy QP commands to ioctl broke the EFA
>> provider in [1].
>> With the new ioctl the 'ibv_resp' part of the response is all zero'd, opposed to
>> the write method.
>
> It is a bug in the efa provider. It is not allowed to touch the
> 'ibv_resp' parts of the structure from a provider.
Thanks Jason, I did not know that.
> Something like this
>
> diff --git a/providers/efa/verbs.c b/providers/efa/verbs.c
> index 5f8e7b800210bb..92bb7432a92a2b 100644
> --- a/providers/efa/verbs.c
> +++ b/providers/efa/verbs.c
> @@ -541,7 +541,9 @@ static void efa_sq_terminate(struct efa_qp *qp)
> efa_wq_terminate(&qp->sq.wq);
> }
>
> -static int efa_sq_initialize(struct efa_qp *qp, struct efa_create_qp_resp *resp)
> +static int efa_sq_initialize(struct efa_qp *qp,
> + const struct ibv_qp_init_attr_ex *attr,
> + struct efa_create_qp_resp *resp)
> {
> struct efa_dev *dev = to_efa_dev(qp->verbs_qp.qp.context->device);
> size_t desc_ring_size;
> @@ -559,7 +561,7 @@ static int efa_sq_initialize(struct efa_qp *qp, struct efa_create_qp_resp *resp)
> desc_ring_size = qp->sq.wq.wqe_cnt * sizeof(struct efa_io_tx_wqe);
> qp->sq.desc_ring_mmap_size = align(desc_ring_size + qp->sq.desc_offset,
> qp->page_size);
> - qp->sq.max_inline_data = resp->ibv_resp.max_inline_data;
> + qp->sq.max_inline_data = attr->cap.max_inline_data;
>
> qp->sq.local_queue = malloc(desc_ring_size);
> if (!qp->sq.local_queue) {
> @@ -840,7 +842,7 @@ static struct ibv_qp *create_qp(struct ibv_context *ibvctx,
> if (err)
> goto err_destroy_qp;
>
> - err = efa_sq_initialize(qp, &resp);
> + err = efa_sq_initialize(qp, attr, &resp);
> if (err)
> goto err_terminate_rq;
>
>
I'll send a PR.
