[openib-general] [PATCH/RFC 1/2] IB: Return "maybe_missed_event" hint from ib_req_notify_cq()

[Shirley Ma]

Hi, Roland,

We have found missing interrupts in ehca driver none scaling code. We are
testing the patch now. I will let you know when we pass the test ASAP.

Does your patch use netif_receive_skb or netif_rx_ni() in IPoIB receiving
path? I haven't looked at your most recent git tree yet. If it's
netif_rx_ni(), that's wrong. NAPI should avoid IP backlog queue.

As we discussed before, I suggested to use return (unlikely(missed_event) &
netif_reschedule_rx()) instead of going back polling cq again and again.
ehca delivers packets too fast, according to my debug output, I could get
up to 58 missed_events between notify_cq and netif_reschedule_rx() to exit
from NAPI poll.

Sorry to block your NAPI patch that long. Are you still planning to use
NAPI as default or as an configuration option? As Michael's pointed out,
under some situation (like heavy load), NAPI might not be a good choice.

Thanks
Shirley Ma
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openfabrics.org/pipermail/general/attachments/20061206/0f926a49/attachment.html>