[openib-general] RE: [PATCH] Opensm - add info in vendor errorprinting

[Hal Rosenstock]

On Thu, 2005-11-17 at 07:49, Yael Kalka wrote:
> Hello all,
>
> During some opensm tests we've encountered a situation where a call
> to umad_send() fails on very large mads (on mads of size ~90,000).
> Is there some limitation on the size of the mads that can be sent?

Those errors are the result of a kmalloc in user_mad.c (ib_umad module)
when sending. If a contiguous block of that size can't be obtained, then
it would fail.

> Reproduction of the issue is simple:
> Run opensm
> Then run osmtest -f m -M 3
> What the osmtest does in this case is try to create as many multicast
> groups as possible (for all the possible MC_lids), and then
> does an SA query on all the multicast groups that exist.
> OpenSM will then try to answer with a huge group of multicast records,
> and when opensm does umad_send() - it fails.
>
> Any ideas why this happens?

Not yet. I now have this on my list but it is not top priority. I am
still digging out from SC05.

-- Hal