[openib-general] [PATCH 0/5 v3] 2.6.20 rdma/cma: add userspace support

[Or Gerlitz]

Roland Dreier wrote:
>  > What's the status of this patchset? it would be somehow very usefull
>  > to have rdma cm user space support enablement in 2.6.20 and without
>  > the multicast code i don't see why not merging it.
> 
> I would like to merge it, but I need to find time to read it over
> carefully.  Have you read this patch set over?  Do you have any
> comments about anything?

+ 1/5 is a small fix discussed over the list

+ 2/5 provides a functionality needed by CMA consumer and does not have 
any impact on anything below the CMA

+ 3/5 is a solution for the IB race of data crossing the RTU and is the 
outcome of a very long discussion over the list. The approach taken is 
very clean and easy to integrate for CM/CMA consumers. A similar patch 
was integrated into OFED 1.1 so it closes a hole where a passive side 
CM/CMA consumers wanting to handle this case easily were not able to do 
so with the kernel CM/CMA code, we must need it for 2.6.20 to close this 
gap.

+ 4/5 adds CMA "UD offload" support using SIDR REQ/REP to exchange the 
QP and Path information. I did not experience much with the patch other 
then running the librdmacm uddady test program but have reviewed it 
without having any special comments.

+ 5/5 is the CMA user space support. I only did a light review of it but 
  my understanding is that Sean used the in kernel ib_ucm design/code as 
the base line for this driver so there should be no special issues here.
This driver is long time missing in the kernel IB offer, as it enables 
using the user space rdma cm (librdmacm) which more and more becomes a 
must have in the IB package of today's distros - it better go in now.

Actually i did most of my review and testing on the multicast code which 
is not in this patch set. I have provided feedback over the list which 
made its way into v2 of the patches and more feedback 1x1 to Sean during 
the sc06 devcon.

Sean - as of the stability issues reported by Mellanox I understand you 
have decided not to push the multicast code for 2.6.20 and I see that 
the focus now is on finding the bug. Once this is solved I would like to 
provide more feedback before you publish v3 - does it makes sense?

Or.