[openib-general] Re: RFC: ib_cache_event problems

[Roland Dreier]

    Michael> How then can we solve the problem of IPoIB querying the
    Michael> cache as a result of an event, and getting a stale value?
    Michael> Note we are actually seeing this in practice when
    Michael> changing pkeys.

It doesn't seem like a severe problem to me -- IPoIB will just check
again in another second, right?

The whole intention of the cache interface is that it should only be
used when a stale value is not fatal.  So if this isn't good enough
for IPoIB, then it should just query the P_Key table directly.  But of
course even that could return stale values, since there's no guarantee
that the P_Key table doesn't change immediately after the query operation.

    Michael> IMO order of events is typically not important, so we
    Michael> only need to handle up to 6 different events in some kind
    Michael> of bitmask.

This seems like a strong statement -- certainly there's a big
difference between the sequence "port active" then "port error"
vs. "port error" then "port active".  Also coalescing events means
that the sequences "port error", "port active", "port error" vs. just
"port error", "port active" can't be distinguished.

So I think the proposed cure may be worse than the disease here.

 - R.