Re: [RFC PATCH] drivers: ata: Read Rx water mark value from device-tree

From: Rob Herring
Date: Tue Feb 23 2016 - 14:29:52 EST


On Tue, Feb 23, 2016 at 03:29:55PM +0000, Anurag Kumar Vulisha wrote:
> Hi Arnd,
>
> > -----Original Message-----
> > From: Arnd Bergmann [mailto:arnd@xxxxxxxx]
> > Sent: Tuesday, February 23, 2016 3:51 PM
> > To: Anurag Kumar Vulisha
> > Cc: robh+dt@xxxxxxxxxx; pawel.moll@xxxxxxx; mark.rutland@xxxxxxx;
> > ijc+devicetree@xxxxxxxxxxxxxx; galak@xxxxxxxxxxxxxx; tj@xxxxxxxxxx;
> > devicetree@xxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx; linux-
> > ide@xxxxxxxxxxxxxxx; Anirudha Sarangi; Srikanth Vemula; Punnaiah Choudary
> > Kalluri
> > Subject: Re: [RFC PATCH] drivers: ata: Read Rx water mark value from device-
> > tree
> >
> > On Tuesday 23 February 2016 05:58:32 Anurag Kumar Vulisha wrote:
> > > >
> > > > I don't know what is appropriate because I have no idea what
> > > > Rxwatermark is good for. Can you try describing why we can't just
> > > > set it to the correct value for everyone automatically?
> > > >
> > >
> > > This RX watermark level sets the minimum number of free locations
> > > within the RX FIFO .When the rx fifo level crosses the programmed
> > > watermark level ,sata controller will transmit HOLDS to the device asking it
> > to wait. This happens when dma reads the rx fifo data slower than the device
> > is sending the data. Note that it can take some time for the HOLDs to get to
> > the other end and in the time there must be enough room in the FIFO to
> > absorb all data that could arrive from the device.
> > > Currently we are using 0x40 for this value, which works fine with all
> > > hardware designs we are currently having. But hoping that this value
> > > may vary for future silicon versions, I wanted to make this as a configurable
> > value. So for this reason I thought of moving it either to device-tree or
> > making it as a module_param() property.
> > >
> >
> > Ok, so if this depends on the silicon version, your initial approach would be
> > better than the module_param.
> >
> > I would probably make this dependent on the compatible string instead, and
> > have a table in the device driver that uses a specific value for each variant of
> > the device, but either way should be fine.
> >
> > Having a separate property is most appropriate if for each hardware revision
> > there is exactly one ideal value, while a table in the driver makes more sense
> > if this takes a bit of tuning and the driver might choose to optimize it
> > differently based on other constraints, such as its own interrupt handler
> > implementation.
> >
>
> Since we are currently having one value in common for all the hardware and also changing
> the rx water mark does not require any changes other than vendor specific PTC register update ,
> I think it would be better to use device tree property for that rx watermark value. Doing
> this makes the updating of rx watermark value easy, if any changes required.
>
> In future, if any silicon version rx water mark value doesn't work with the current versions,
> then I will do as you said by maintaining the table in device driver. But at present I feel
> that single rx watermark property in device tree would be enough, since it works with all the
> hardware versions we have.

If you currently have no reason to modify it now, then add it later when
you actually have a use case.

Rob