Re: [PATCH 2/2] IB/core: Set width to 1X for invalid active widths when port is down

[Hal Rosenstock <hal@dev.mellanox.co.il>]

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 <honli@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 <honli@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.

Thanks.

-- Hal