Re: [Suspected Spam] Re: [PATCH for-rc] RDMA/vmw_pvrdma: Fix the active_speed and phys_state value

[Leon Romanovsky <leon@kernel.org>]

On Tue, Nov 03, 2020 at 02:41:06PM -0800, Adit Ranadive wrote:
> On 11/2/20 10:56 PM, Leon Romanovsky wrote:
> > On Mon, Nov 02, 2020 at 02:46:40PM -0400, Jason Gunthorpe wrote:
> >> On Mon, Nov 02, 2020 at 10:38:19AM -0800, Adit Ranadive wrote:
> >>> On 11/2/20 10:27 AM, Jason Gunthorpe wrote:
> >>>> On Mon, Nov 02, 2020 at 10:21:21AM -0800, Adit Ranadive wrote:
> >>>>> On 11/2/20 10:02 AM, Jason Gunthorpe wrote:
> >>>>>> On Mon, Nov 02, 2020 at 09:55:25AM -0800, Adit Ranadive wrote:
> >>>>>>> On 10/29/20 9:16 AM, Adit Ranadive wrote:
> >>>>>>>> On 10/29/20 4:57 AM, Jason Gunthorpe wrote:
> >>>>>>>>> On Wed, Oct 28, 2020 at 11:19:45PM +0000, Adit Ranadive wrote:
> >>>>>>>>>> The PVRDMA device still reports the active_speed in u8.
> >>>>>>>>>> Lets use the ib_eth_get_speed to report the speed and
> >>>>>>>>>> width. Unfortunately, phys_state gets stored as msb of
> >>>>>>>>>> the new u16 active_speed.
> >>>>>>>>>
> >>>>>>>>> This explanation is not clear, I have no idea what this is fixing
> >>>>>>>>
> >>>>>>>> It seemed more clear to me in my head, I guess :).
> >>>>>>>>
> >>>>>>>> After commit 376ceb31ff87 changed the active_speed attribute to
> >>>>>>>> u16, both the active_speed and phys_state attributes in the
> >>>>>>>> pvrdma_port_attr struct are getting stored in this u16. As a
> >>>>>>>> result, these show up as invalid values in ibv_devinfo.
> >>>>>>>>
> >>>>>>>> Our device still gives us back a u8 active_speed so both these
> >>>>>>>> are getting stored in the u16. This fix I proposed simply gets
> >>>>>>>> the active_speed from the netdev while the phys_state still
> >>>>>>>> needs to come from the pvrdma device, i.e. the msb the of the
> >>>>>>>> u16. I also removed some unused functions as a result.
> >>>>>>>>
> >>>>>>>> Alternatively, I could change the u8 active_width and u16
> >>>>>>>> active_speed to reserved now that we're getting the active_speed
> >>>>>>>> and active_width from the ib_get_eth_speed function.
> >>>>>>>
> >>>>>>> Jason, did you have any comments on this or did you want me
> >>>>>>> to just send v1 with an updated description?
> >>>>>>
> >>>>>> I still haven't figured out what this is fixing.
> >>>>>>
> >>>>>> Is 'struct pvrdma_port_attr' some kind of ABI? If so why isn't the fix
> >>>>>> to revert the type?
> >>>>>
> >>>>> I can revert it but I thought that it had to a u16 based on the IBTA, no?
> >>>>> Or does that not apply to device-level stuff?
> >>>>
> >>>> You didn't answer the question, it it ABI to some kind of FW interface
> >>>> or something?
> >>>>
> >>>> *HOW* did two fields get overlapped onto a single u16?? The compiler
> >>>> won't do this..
> >>>>
> >>>
> >>> It is an ABI to the device for port attributes. The device gives us back
> >>> this structure for query port verb. The response from the device is
> >>> memcopied into this pvrdma_port_attr structure. So both the bytes
> >>> representing active_speed and phys_state from the device are copied
> >>> into the single u16 in this structure.
> >>
> >> So it is ABI and it shouldn't have been changed, point at the stuff
> >> that made it ABI and revert the structure layout change..
> >
> > How will it work for the new IBTA speed?
>
> Hopefully that should be addressed in another patch I'll send out
> that uses the ib_get_eth_speed api?

It will work for now, but any new speed above NDR will break pvrdma.

Thanks

>
> >
> > Thanks
> >
> >>
> >> Jason