Re: [PATCH rdma-rext 0/4] RDMA/bnxt_re: Add QP rate limit support

[Leon Romanovsky <leon@kernel.org>]

On Wed, Jan 28, 2026 at 12:33:57PM +0530, Selvin Xavier wrote:
> On Tue, Jan 27, 2026 at 1:49 AM Leon Romanovsky <leon@kernel.org> wrote:
> >
> > On Sun, Jan 25, 2026 at 09:47:05AM +0530, Kalesh Anakkur Purayil wrote:
> > > Hi Leon, Jason,
> > >
> > > A gentle reminder. Could you please review the patch series?
> >
> > Sorry for the delayed response. The idea and implementation look fine to me.
> Thanks you for your response.
> > What is missing is a clear and well-documented definition of the semantics
> > of QP rate limiting for RC QPs.
> 
> man page of ibv_modify_qp doesn't have much information about rate limit
> 
>  struct ibv_qp_attr {
> ...
>           uint32_t                rate_limit;             /* Rate
> limit in kbps for packet pacing */
>        };
> 
> 
> attr_mask:IBV_QP_RATE_LIMIT  Set rate_limit
> 
> This man page contains only the required field for each transition and
> doesn't mention about the optional flags. Do you want us to add a
> section for the QP rate limit in the notes?

I would expect the changes to be in https://github.com/linux-rdma/rdma-core/blob/master/libibverbs/man/ibv_modify_qp_rate_limit.3
and your driver implements .modify_qp_rate_limit() callback in rdma-core.

> 
> >
> > How should RDMA_READ or small RDMA_REQ packets be treated? Are response
> > packets included in the rate limit as well? It must be documented in
> > man pages.
> 
> All transmitted packets (including rdma_read request and other request
> packets) will be part of rate limiting setting for the QP. In our
> implementation, the ack packets are not part of the rate limiting of
> the normal transmit path. READ data response will be rate limited at
> the peer side, if the rate limit is configured on the peer side.
> 
> We have another question. Existing implementation of IB_QP_RATE_LIMIT
> is applicable only for raw ethernet QP. With this change, we will
> start supporting for RC QPs also. So mlx driver can also get this
> request for RC QPs, but it will silently ignored as the QP type is not
> Raw ethernet QP. Should we fail the request instead?

We should return something to users, since it would be surprising if a
policy-related feature simply failed without notice.

The tricky part is deciding how to expose this so users can anticipate it
up front, especially since we're dealing with a generic verb.

Thanks

> 
> Thanks,
> Selvin Xavier
> >
> > Thanks
> >
> > >
> > > On Fri, Jan 16, 2026 at 2:43 PM Kalesh AP
> > > <kalesh-anakkur.purayil@broadcom.com> wrote:
> > > >
> > > > Hi,
> > > >
> > > > This patchset supports QP rate limit in the bnxt_re driver.
> > > >
> > > > Broadcom P7 devices supports setting the rate limit while changing
> > > > RC QP state from INIT to RTR, RTR to RTS and RTS to RTS. Or, once
> > > > the QP is transitioned to RTR or RTS state.
> > > >
> > > > First patch adds stack support for rate limit for RC QPs.
> > > >
> > > > Second patch adds support for QP rate limiting in the bnxt_re driver.
> > > >
> > > > Third patch adds support to report packet pacing capabilities in the
> > > > query_device.
> > > >
> > > > Forth patch adds support to report QP rate limit in debugfs QP info.
> > > >
> > > > The pull request for rdma-core changes are at:
> > > >
> > > > https://github.com/linux-rdma/rdma-core/pull/1692
> > > >
> > > > Regards,
> > > > Kalesh
> > > >
> > > > Kalesh AP (4):
> > > >   IB/core: Extend rate limit support for RC QPs
> > > >   RDMA/bnxt_re: Add support for QP rate limiting
> > > >   RDMA/bnxt_re: Report packet pacing capabilities when querying device
> > > >   RDMA/bnxt_re: Report QP rate limit in debugfs
> > > >
> > > >  drivers/infiniband/core/verbs.c           |  9 ++++--
> > > >  drivers/infiniband/hw/bnxt_re/debugfs.c   | 14 ++++++--
> > > >  drivers/infiniband/hw/bnxt_re/ib_verbs.c  | 39 +++++++++++++++++++++--
> > > >  drivers/infiniband/hw/bnxt_re/qplib_fp.c  | 12 ++++++-
> > > >  drivers/infiniband/hw/bnxt_re/qplib_fp.h  |  3 ++
> > > >  drivers/infiniband/hw/bnxt_re/qplib_res.h |  6 ++++
> > > >  drivers/infiniband/hw/bnxt_re/qplib_sp.c  |  5 +++
> > > >  drivers/infiniband/hw/bnxt_re/qplib_sp.h  |  2 ++
> > > >  drivers/infiniband/hw/bnxt_re/roce_hsi.h  | 13 +++++---
> > > >  include/uapi/rdma/bnxt_re-abi.h           | 16 ++++++++++
> > > >  10 files changed, 107 insertions(+), 12 deletions(-)
> > > >
> > > > --
> > > > 2.43.5
> > > >
> > >
> > >
> > > --
> > > Regards,
> > > Kalesh AP
> >
> >
> >