On 01/07/2022 03:04, Bob Pearson wrote:
> The code thc that decides whether to defer execution of a wqe in
> rxe_requester.c is isolated into a subroutine rxe_is_fenced()
> and removed from the call to req_next_wqe(). The condition whether
> a wqe should be fenced is changed to comply with the IBA. Currently
> an operation is fenced if the fence bit is set in the wqe flags and
> the last wqe has not completed. For normal operations the IBA
> actually only requires that the last read or atomic operation is
> complete.
>
> Signed-off-by: Bob Pearson <rpearsonhpe@xxxxxxxxx>
Looks good to me.
Reviewed-by: Li Zhijian <lizhijian@xxxxxxxxxxx>
> ---
> v2 replaces "RDMA/rxe: Fix incorrect fencing"
>
> drivers/infiniband/sw/rxe/rxe_req.c | 37 ++++++++++++++++++++++++-----
> 1 file changed, 31 insertions(+), 6 deletions(-)
>
> diff --git a/drivers/infiniband/sw/rxe/rxe_req.c b/drivers/infiniband/sw/rxe/rxe_req.c
> index 9d98237389cf..e8a1664a40eb 100644
> --- a/drivers/infiniband/sw/rxe/rxe_req.c
> +++ b/drivers/infiniband/sw/rxe/rxe_req.c
> @@ -161,16 +161,36 @@ static struct rxe_send_wqe *req_next_wqe(struct rxe_qp *qp)
> (wqe->state != wqe_state_processing)))
> return NULL;
>
> - if (unlikely((wqe->wr.send_flags & IB_SEND_FENCE) &&
> - (index != cons))) {
> - qp->req.wait_fence = 1;
> - return NULL;
> - }
> -
> wqe->mask = wr_opcode_mask(wqe->wr.opcode, qp);
> return wqe;
> }
>
> +/**
> + * rxe_wqe_is_fenced - check if next wqe is fenced
> + * @qp: the queue pair
> + * @wqe: the next wqe
> + *
> + * Returns: 1 if wqe needs to wait
> + * 0 if wqe is ready to go
> + */
> +static int rxe_wqe_is_fenced(struct rxe_qp *qp, struct rxe_send_wqe *wqe)
> +{
> + /* Local invalidate fence (LIF) see IBA 10.6.5.1
> + * Requires ALL previous operations on the send queue
> + * are complete. Make mandatory for the rxe driver.
> + */
> + if (wqe->wr.opcode == IB_WR_LOCAL_INV)
> + return qp->req.wqe_index != queue_get_consumer(qp->sq.queue,
> + QUEUE_TYPE_FROM_CLIENT);
> +
> + /* Fence see IBA 10.8.3.3
> + * Requires that all previous read and atomic operations
> + * are complete.
> + */
> + return (wqe->wr.send_flags & IB_SEND_FENCE) &&
> + atomic_read(&qp->req.rd_atomic) != qp->attr.max_rd_atomic;
> +}
> +
> static int next_opcode_rc(struct rxe_qp *qp, u32 opcode, int fits)
> {
> switch (opcode) {
> @@ -632,6 +652,11 @@ int rxe_requester(void *arg)
> if (unlikely(!wqe))
> goto exit;
>
> + if (rxe_wqe_is_fenced(qp, wqe)) {
> + qp->req.wait_fence = 1;
> + goto exit;
> + }
> +
> if (wqe->mask & WR_LOCAL_OP_MASK) {
> ret = rxe_do_local_ops(qp, wqe);
> if (unlikely(ret))
