Re: [PATCH rdma-next 26/50] RDMA/erdma: Separate user and kernel CQ creation paths

[Leon Romanovsky <leon@kernel.org>]

On Thu, Feb 26, 2026 at 02:17:38PM +0800, Junxian Huang wrote:
> 
> 
> On 2026/2/13 18:58, Leon Romanovsky wrote:
> > From: Leon Romanovsky <leonro@nvidia.com>
> > 
> > Split CQ creation into distinct kernel and user flows. The hns driver,
> > inherited from mlx4, uses a problematic pattern that shares and caches
> > umem in hns_roce_db_map_user(). This design blocks the driver from
> > supporting generic umem sources (VMA, dmabuf, memfd, and others).
> > 
> > In addition, let's delete counter that counts CQ creation errors. There
> > are multiple ways to debug kernel in modern kernel without need to rely
> > on that debugfs counter.
> > 
> > Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
> > ---
> >  drivers/infiniband/hw/hns/hns_roce_cq.c      | 103 ++++++++++++++++++++-------
> >  drivers/infiniband/hw/hns/hns_roce_debugfs.c |   1 -
> >  drivers/infiniband/hw/hns/hns_roce_device.h  |   3 +-
> >  drivers/infiniband/hw/hns/hns_roce_main.c    |   1 +
> >  4 files changed, 82 insertions(+), 26 deletions(-)

<...>

> > +int hns_roce_create_cq(struct ib_cq *ib_cq, const struct ib_cq_init_attr *attr,
> > +		       struct uverbs_attr_bundle *attrs)
> > +{
> > +	struct hns_roce_dev *hr_dev = to_hr_dev(ib_cq->device);
> > +	struct hns_roce_ib_create_cq_resp resp = {};
> > +	struct hns_roce_cq *hr_cq = to_hr_cq(ib_cq);
> > +	struct ib_device *ibdev = &hr_dev->ib_dev;
> > +	struct hns_roce_ib_create_cq ucmd = {};
> 
> ucmd and resp are not needed since we don't have udata here.

Thanks, will fix.

> 
> Junxian