[ofa-general] [Bug 508] IPoIB CM multicast is hogging interrupts

[Pradeep Satyanarayana]

> I think this trick I just came up with is a simpe way to prevent
> IPoIB TX from hogging interrupts, even without NAPI. And it might be a 
better
> way to solve the problem for IPoIB CM TX than using a common cq
> as my previous patch did.
> 
> This seems to hurt top bandwidth a bit in my testing, so this needs some 
more
> work. Meanwhile, Scott, could you please check whether the following
> patch helps in your test-case?
> 
> Roland, I think something similiar is a good idea for SRP, too.
> What do you think?
> 
> diff --git a/drivers/infiniband/ulp/ipoib/ipoib_cm.c 
> b/drivers/infiniband/ulp/ipoib/ipoib_cm.c
> index 2b242a4..3ed1536 100644
> --- a/drivers/infiniband/ulp/ipoib/ipoib_cm.c
> +++ b/drivers/infiniband/ulp/ipoib/ipoib_cm.c
> @@ -573,14 +573,15 @@ static void ipoib_cm_handle_tx_wc(struct 
> net_device *dev, struct ipoib_cm_tx *tx
>  static void ipoib_cm_tx_completion(struct ib_cq *cq, void *tx_ptr)
>  {
>     struct ipoib_cm_tx *tx = tx_ptr;
> -   int n, i;
> +   int n, i, cnt = 0;
> 
>     ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
>     do {
>        n = ib_poll_cq(cq, IPOIB_NUM_WC, tx->ibwc);
> +      cnt += n;
>        for (i = 0; i < n; ++i)
>           ipoib_cm_handle_tx_wc(tx->dev, tx, tx->ibwc + i);
> -   } while (n == IPOIB_NUM_WC);
> +   } while (n == IPOIB_NUM_WC && cnt < ipoib_sendq_size);
>  }

This change might exit tx_completion sooner -how does that prevent
hogging interrupts (without NAPI)? I am not clear about that.

When NAPI is merged, would this be equivalent of running out of budget
and treated accordingly? Otherwise, one might continue to be in polling 
mode
for a longer period and maybe starving other interfaces.

Pradeep
pradeep at us.ibm.com