Re: [PATCH rdma-next 1/1] RDMA/mana_ib: process QP error events

From: Leon Romanovsky
Date: Tue Jun 04 2024 - 08:08:58 EST


On Tue, Jun 04, 2024 at 04:36:03AM -0700, Konstantin Taranov wrote:
> From: Konstantin Taranov <kotaranov@xxxxxxxxxxxxx>
>
> Process QP fatal events from the error event queue.
> For that, find the QP, using QPN from the event, and then call its
> event_handler. To find the QPs, store created RC QPs in an xarray.
>
> Signed-off-by: Konstantin Taranov <kotaranov@xxxxxxxxxxxxx>
> Reviewed-by: Wei Hu <weh@xxxxxxxxxxxxx>
> ---
> drivers/infiniband/hw/mana/device.c | 3 ++
> drivers/infiniband/hw/mana/main.c | 37 ++++++++++++++++++-
> drivers/infiniband/hw/mana/mana_ib.h | 4 ++
> drivers/infiniband/hw/mana/qp.c | 11 ++++++
> .../net/ethernet/microsoft/mana/gdma_main.c | 1 +
> include/net/mana/gdma.h | 1 +
> 6 files changed, 55 insertions(+), 2 deletions(-)

<...>

> +static void
> +mana_ib_event_handler(void *ctx, struct gdma_queue *q, struct gdma_event *event)
> +{
> + struct mana_ib_dev *mdev = (struct mana_ib_dev *)ctx;
> + struct mana_ib_qp *qp;
> + struct ib_event ev;
> + unsigned long flag;
> + u32 qpn;
> +
> + switch (event->type) {
> + case GDMA_EQE_RNIC_QP_FATAL:
> + qpn = event->details[0];
> + xa_lock_irqsave(&mdev->qp_table_rq, flag);
> + qp = xa_load(&mdev->qp_table_rq, qpn);
> + if (qp)
> + refcount_inc(&qp->refcount);
> + xa_unlock_irqrestore(&mdev->qp_table_rq, flag);
> + if (!qp)
> + break;
> + if (qp->ibqp.event_handler) {
> + ev.device = qp->ibqp.device;
> + ev.element.qp = &qp->ibqp;
> + ev.event = IB_EVENT_QP_FATAL;
> + qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
> + }
> + if (refcount_dec_and_test(&qp->refcount))
> + complete(&qp->free);
> + break;
> + default:
> + break;
> + }
> +}

<...>

> @@ -620,6 +626,11 @@ static int mana_ib_destroy_rc_qp(struct mana_ib_qp *qp, struct ib_udata *udata)
> container_of(qp->ibqp.device, struct mana_ib_dev, ib_dev);
> int i;
>
> + xa_erase_irq(&mdev->qp_table_rq, qp->ibqp.qp_num);
> + if (refcount_dec_and_test(&qp->refcount))
> + complete(&qp->free);
> + wait_for_completion(&qp->free);

This flow is unclear to me. You are destroying the QP and need to make
sure that mana_ib_event_handler is not running at that point and not
mess with refcount and complete.

Thanks