On Thu, Mar 15, 2018 at 08:47:44AM -0400, Hal Rosenstock wrote:
> On 3/15/2018 8:43 AM, Honggang LI wrote:
> > On Thu, Mar 15, 2018 at 08:32:02AM -0400, Hal Rosenstock wrote:
> >> On 3/15/2018 8:01 AM, Hal Rosenstock wrote:
> >>> On 3/15/2018 5:02 AM, Honggang LI wrote:
> >>>> From: Honggang Li <ho...@redhat.com>
> >>>>
> >>>> commit f1b65df5a232 ("IB/mlx5: Add support for active_width and
> >>>> active_speed in RoCE"). Before this patch applied, the mlx5_ib
> >>>> driver set default active_width and active_speed to IB_WIDTH_4X
> >>>> and IB_SPEED_QDR.
> >>>>
> >>>> When the RoCE port is down, the RoCE port did not negotiate the
> >>>> active width with remote side. The active width is zero. If run
> >>>> ibstat to require the port status, ibstat will panic as it read
> >>>> invalid width from sys file.
> >>>>
> >>>> Signed-off-by: Honggang Li <ho...@redhat.com>
> >>>> ---
> >>>> drivers/infiniband/core/sysfs.c | 15 +++++++++++----
> >>>> 1 file changed, 11 insertions(+), 4 deletions(-)
> >>>>
> >>>> diff --git a/drivers/infiniband/core/sysfs.c
> >>>> b/drivers/infiniband/core/sysfs.c
> >>>> index cf36ff1f0068..722e4571f4d2 100644
> >>>> --- a/drivers/infiniband/core/sysfs.c
> >>>> +++ b/drivers/infiniband/core/sysfs.c
> >>>> @@ -240,6 +240,7 @@ static ssize_t rate_show(struct ib_port *p, struct
> >>>> port_attribute *unused,
> >>>> struct ib_port_attr attr;
> >>>> char *speed = "";
> >>>> int rate; /* in deci-Gb/sec */
> >>>> + int width;
> >>>> ssize_t ret;
> >>>>
> >>>> ret = ib_query_port(p->ibdev, p->port_num, &attr);
> >>>> @@ -278,13 +279,19 @@ static ssize_t rate_show(struct ib_port *p, struct
> >>>> port_attribute *unused,
> >>>> break;
> >>>> }
> >>>>
> >>>> - rate *= ib_width_enum_to_int(attr.active_width);
> >>>> - if (rate < 0)
> >>>> - return -EINVAL;
> >>>> + width = ib_width_enum_to_int(attr.active_width);
> >>>> + if (width < 0) {
> >>>> + if (attr.state != IB_PORT_ACTIVE)
> >>>
> >>> Link width is valid in any PortState other than Down so I think that
> >>> this check should be:
> >>> if (attr.state != IB_PORT_DOWN)
> >>>
> >>> However, I don't think overriding width should be needed for this case
> >>> and just returning -EINVAL should be fine regardless of port state.
> >>> AFAIK it's the driver responsibility to populate acceptable defaults for
> >>> such parameters. What driver(s) have this issue ? Shouldn't it be fixed
> >>> there rather than here ?
> >>
> >> I just noticed that you reference commit f1b65df5a232 ("IB/mlx5: Add
> >> support for active_width and active_speed in RoCE"). Before this patch
> >> applied, the mlx5_ib driver set default active_width and active_speed to
> >> IB_WIDTH_4X and IB_SPEED_QDR.
> >>
> >> Should the fix be to hw/mlx5/main.c:translate_eth_proto_oper which now has:
> >>
> >> switch (eth_proto_oper) {
> >> ...
> >> default:
> >> return -EINVAL;
> >> }
> >>
> >> return 0;
> >>
> >> and change default case to:
> >> *active_width = IB_WIDTH_1X;
> >>
> >> ?
> >
> > I suggest to restore previous behavior before apply f1b65df5a232.
> >
> > diff --git a/drivers/infiniband/hw/mlx5/main.c
> > b/drivers/infiniband/hw/mlx5/main.c
> > index 033b6af90de9..0d73d2772d9b 100644
> > --- a/drivers/infiniband/hw/mlx5/main.c
> > +++ b/drivers/infiniband/hw/mlx5/main.c
> > @@ -388,6 +388,9 @@ static int mlx5_query_port_roce(struct ib_device
> > *device, u8 port_num,
> > if (err)
> > goto out;
> >
> > + props->active_width = IB_WIDTH_4X;
> > + props->active_speed = IB_SPEED_QDR;
> > +
> > translate_eth_proto_oper(eth_prot_oper, &props->active_speed,
> > &props->active_width);
> >
>
> Sure; makes sense that it should preserve the original behavior for this
> case.
[PATCH] IB/mlx5: Set the default active rate and width to QDR and 4X
This new patch had been sent to upstream mailing list for review.
thanks
