[ofa-general] Re: [GIT PULL] please pull infiniband.git

[Michael S. Tsirkin]

>  > If stale CQEs could come out, you would never be able to
>  > destroy a QP without destroying the CQ.
> 
> Why?

Assume you stick a pointer in WR_ID. When is it safe to free
the object?

But if you look at the language in 10.2.4.4, you actually
are not *supposed* to destroy a QP that has outstanding WRs.

So for send side, it now seems the bug is in IPoIB - it should
be fixed to drain the send queue rather than trying to destroy
QP directly.

However, it seems IB spec has a hole - the procedure outlined there
can not work for SRQ.

-- 
MST