[ofa-general] Re: [PATCH v2] IB/ipoib: Split CQs for IPOIB UD

[Roland Dreier]

Umm... a little late now that I asked Linus to pull but I realized that
this patch is somewhat buggy by design:

You make the send CQ unsignaled, so you never get TX completion events.
But this means that if the send queue ever fills up, we'll do
netif_stop_queue() and then never reap a TX completion to restart the
queue... I guess if we do netif_stop_queue() then we had better start a
timer or something to kick us sometime in the future.

Or we could request an event for the send CQ only when the send queue
is full.

But polling from either a timer or CQ event leads to locking issues
against polling from the send path...

 - R.