Jason Gunthorpe

The information in libibmad does not seem to be complete. What is the attribute ID of the vendor structure, what operation is used to read it? The layout in libibmad...

It is hand written from the specification of the wire format. I expect that specific mad is out of date, but something that looks basically like it is what is...

It is very interesting to see that this is even possible, but to go forward I think a little more finesse would be needed. There is room in the API...

Thank you for the patch, What scenarios were you able to use this in? There are many reasons a SMP send during discovery could fail, this seems to deal with...

This seems like a good improvement until a generally more complete way of handling discovery errors could be implemented. What I had in mind was removing all the failed links...

On Tue, Jul 10, 2012 at 07:39:36AM -0700, Christian Kniep wrote: > we maintain a fairly large setup with a broad range of nodes in > which there are nodes...

protocol not supported means lack of kernel support for the MLX5_IB_METHOD_DEVX_ASYNC_EVENT_FD_ALLOC This was added on more recently, it wasn't in the first devx release.

No, try and fail. It is a fairly easy try and fail, just call it and immediately close the fd after.

No UD is fine, providers just have to provide some level of RDMA capability, like RC is fine. iWarp for instance doesn't support UD. I also know nothing about qedr...

mips may have memory barriers, but do any DMA cache coherent MIPS implementations exist? Which ones are they?