Re: [PATCH for-next 2/2] RDMA/erdma: Add drain_sq and drain_rq support

[Cheng Xu <chengyou@linux.alibaba.com>]

On 8/24/22 5:42 PM, Cheng Xu wrote:
> For erdma, hardware won't process any WRs after modifying QP state to
> error, so the default __ib_drain_sq and __ib_drain_rq can not work for
> erdma device. Here, we introduce custom implementation of drain_sq and
> drain_rq interface to fit erdma hardware.
> 
> Signed-off-by: Cheng Xu <chengyou@linux.alibaba.com>
> ---
>  drivers/infiniband/hw/erdma/erdma_qp.c    | 71 +++++++++++++++++++++++
>  drivers/infiniband/hw/erdma/erdma_verbs.h | 10 ++++
>  2 files changed, 81 insertions(+)
> 
> diff --git a/drivers/infiniband/hw/erdma/erdma_qp.c b/drivers/infiniband/hw/erdma/erdma_qp.c
> index abf8b134d076..57fdb946fbfd 100644
> --- a/drivers/infiniband/hw/erdma/erdma_qp.c
> +++ b/drivers/infiniband/hw/erdma/erdma_qp.c
> @@ -599,3 +599,74 @@ int erdma_post_recv_nodrain(struct ib_qp *ibqp,
>  {
>  	return erdma_post_recv(ibqp, recv_wr, bad_recv_wr, false);
>  }
> +
> +static void erdma_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
> +{
> +	struct erdma_drain_cqe *cqe =
> +		container_of(wc->wr_cqe, struct erdma_drain_cqe, cqe);
> +
> +	complete(&cqe->done);
> +}
> +
> +static void erdma_drain_qp_common(struct ib_qp *ibqp, struct completion *comp,
> +				  struct ib_cq *ibcq)
> +{
> +	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
> +	struct erdma_qp *qp = to_eqp(ibqp);
> +	const struct ib_send_wr *bad_swr;
> +	const struct ib_recv_wr *bad_rwr;
> +	struct ib_rdma_wr swr = {
> +		.wr = {
> +			.next = NULL,
> +			{ .wr_cqe   = &qp->kern_qp.sdrain.cqe, },
> +			.opcode = IB_WR_RDMA_WRITE,
> +			.send_flags = IB_SEND_SIGNALED,
> +		},
> +	};
> +	struct ib_recv_wr rwr = {
> +		.next = NULL,
> +		.wr_cqe = &qp->kern_qp.rdrain.cqe,
> +		.num_sge = 0,
> +	};
> +
> +	if (qp->flags & ERDMA_QP_FLAGS_DRAIN_ISSUED)
> +		goto wait_for_completion;
> +
> +	qp->flags |= ERDMA_QP_FLAGS_DRAIN_ISSUED;
> +
> +	qp->kern_qp.rdrain.cqe.done = erdma_drain_qp_done;
> +	init_completion(&qp->kern_qp.rdrain.done);
> +
> +	qp->kern_qp.sdrain.cqe.done = erdma_drain_qp_done;
> +	init_completion(&qp->kern_qp.sdrain.done);
> +
> +	if (erdma_post_recv(ibqp, &rwr, &bad_rwr, true))
> +		return;
> +
> +	if (erdma_post_send(ibqp, &swr.wr, &bad_swr, true))
> +		return;
> +
> +	if (ib_modify_qp(ibqp, &attr, IB_QP_STATE))
> +		return;
> +
> +wait_for_completion:
> +	if (ibcq->poll_ctx == IB_POLL_DIRECT)
> +		while (wait_for_completion_timeout(comp, HZ / 10) <= 0)
> +			ib_process_cq_direct(ibcq, -1);
> +	else
> +		wait_for_completion(comp);
> +}
> +
> +void erdma_drain_sq(struct ib_qp *ibqp)
> +{
> +	struct erdma_qp *qp = to_eqp(ibqp);
> +
> +	erdma_drain_qp_common(ibqp, &qp->kern_qp.sdrain.done, ibqp->send_cq);
> +}
> +
> +void erdma_drain_rq(struct ib_qp *ibqp)
> +{
> +	struct erdma_qp *qp = to_eqp(ibqp);
> +
> +	erdma_drain_qp_common(ibqp, &qp->kern_qp.rdrain.done, ibqp->recv_cq);
> +}
> diff --git a/drivers/infiniband/hw/erdma/erdma_verbs.h b/drivers/infiniband/hw/erdma/erdma_verbs.h
> index f4148fbac878..4cec92c8a737 100644
> --- a/drivers/infiniband/hw/erdma/erdma_verbs.h
> +++ b/drivers/infiniband/hw/erdma/erdma_verbs.h
> @@ -133,6 +133,11 @@ struct erdma_uqp {
>  	u32 rq_offset;
>  };
>  
> +struct erdma_drain_cqe {
> +	struct ib_cqe cqe;
> +	struct completion done;
> +};
> +
>  struct erdma_kqp {
>  	u16 sq_pi;
>  	u16 sq_ci;
> @@ -155,6 +160,9 @@ struct erdma_kqp {
>  	void *sq_db_info;
>  	void *rq_db_info;
>  
> +	struct erdma_drain_cqe sdrain;
> +	struct erdma_drain_cqe rdrain;
> +
>  	u8 sig_all;
>  };
>  
> @@ -341,6 +349,8 @@ int erdma_post_send_nodrain(struct ib_qp *ibqp,
>  int erdma_post_recv_nodrain(struct ib_qp *ibqp,
>  			    const struct ib_recv_wr *recv_wr,
>  			    const struct ib_recv_wr **bad_recv_wr);
> +void erdma_drain_sq(struct ib_qp *ibqp);
> +void erdma_drain_rq(struct ib_qp *ibqp);
>  int erdma_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc);
>  struct ib_mr *erdma_ib_alloc_mr(struct ib_pd *ibpd, enum ib_mr_type mr_type,
>  				u32 max_num_sg);

I'm sorry that I forgot to involved the corresponding assignment of struct ib_device_ops
into this patch. I will send v2 to fix this.


diff --git a/drivers/infiniband/hw/erdma/erdma_main.c b/drivers/infiniband/hw/erdma/erdma_main.c
index 4921ebc1286d..e4ce77607f10 100644
--- a/drivers/infiniband/hw/erdma/erdma_main.c
+++ b/drivers/infiniband/hw/erdma/erdma_main.c
@@ -446,6 +446,8 @@ static const struct ib_device_ops erdma_device_ops = {
        .dereg_mr = erdma_dereg_mr,
        .destroy_cq = erdma_destroy_cq,
        .destroy_qp = erdma_destroy_qp,
+       .drain_rq = erdma_drain_rq,
+       .drain_sq = erdma_drain_sq,
        .get_dma_mr = erdma_get_dma_mr,
        .get_port_immutable = erdma_get_port_immutable,
        .iw_accept = erdma_accept,