[openib-general] ib_cancel_mad and the CM

[Sean Hefty]

Roland Dreier wrote:

> Using wr_ids to hold pointers often seems to lead to issues like this.
> So it might be best to rework the CM so uses wr_id some other way.
> 
> I'll also remind you of the following Sean Hefty quote, circa
> September 2004 ;)
> 
>     "But, yes, users wanting to be able to cancel MADs cannot use duplicate wr_ids."

There are never duplicated values for wr_ids outstanding at the same 
time, which is what the above quote was referring to.  To me, the 
problem seems more related to the inability to synchronize canceling a 
MAD with processing its completion.  Without that ability, it seems 
that the user needs to have some sort of increasing wr_id value.

I did look at what the sa_query did for canceling queries, but I 
_think_ it may have a similar issue.  It's just harding to hit because 
it checks both a pointer and an index.  But if a problem does exist, 
then a user could cancel a query issued by another user.

- Sean