On Fri, Feb 22, 2019 at 4:35 AM Jason Gunthorpe <jgg@xxxxxxxx> wrote:
>
> On Thu, Feb 21, 2019 at 04:13:28AM -0500, Devesh Sharma wrote:
> > In the backdrop of ODP changes recently being done in the
>
> The changes were not related to ODP
Oh! I thought this was a pre-cursor to ODP changes. I will change anyways.
>
> > kernel ib stack and drivers, there was a regression added
> > in the __alloc_pbl path. The change left bnxt_re in DOA
> > state in for-next branch.
> >
> > Fixing the regression to avoid the host crash when a user
> > space object is created. Restricting the unconditional
> > access to hwq.pg_arr when hwq is initialized for a user
> > space object.
>
> .. kernel space object? Isn't it?
Kernel object takes different path and do not traverse the dma-sg-list.
>
> > Fixes: commit 161ebe2498d4 ("RDMA/bnxt_re:
> > Use for_each_sg_dma_page iterator on umem SGL")
> >
> > Signed-off-by: Devesh Sharma <devesh.sharma@xxxxxxxxxxxx>
> > drivers/infiniband/hw/bnxt_re/ib_verbs.c | 13 +++++++++----
> > drivers/infiniband/hw/bnxt_re/qplib_fp.c | 20 ++++++--------------
> > drivers/infiniband/hw/bnxt_re/qplib_res.c | 5 +----
> > 3 files changed, 16 insertions(+), 22 deletions(-)
> >
> > diff --git a/drivers/infiniband/hw/bnxt_re/ib_verbs.c b/drivers/infiniband/hw/bnxt_re/ib_verbs.c
> > index 2ed7786..6150a2f 100644
> > +++ b/drivers/infiniband/hw/bnxt_re/ib_verbs.c
> > @@ -793,9 +793,11 @@ int bnxt_re_destroy_qp(struct ib_qp *ib_qp)
> > {
> > struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp);
> > struct bnxt_re_dev *rdev = qp->rdev;
> > - int rc;
> > + struct ib_pd *ibpd;
> > unsigned int flags;
> > + int rc;
> >
> > + ibpd = qp->ib_qp.pd;
> > bnxt_qplib_flush_cqn_wq(&qp->qplib_qp);
> > rc = bnxt_qplib_destroy_qp(&rdev->qplib_res, &qp->qplib_qp);
> > if (rc) {
> > @@ -803,9 +805,12 @@ int bnxt_re_destroy_qp(struct ib_qp *ib_qp)
> > return rc;
> > }
> >
> > - flags = bnxt_re_lock_cqs(qp);
> > - bnxt_qplib_clean_qp(&qp->qplib_qp);
> > - bnxt_re_unlock_cqs(qp, flags);
> > + if (!ibpd->uobject) {
>
> Looks like the qp needs to be tested for 'userness', not the PD. Use
>
> if (!rdma_is_kernel_res(&qp->res))
>
Sure.
> Jason
