[PATCH 1/2] powerpc/qman: Change fsl,qman-channel-id to cell-index

[Roy Pledge]

> -----Original Message-----
> From: Wood Scott-B07421
> Sent: Wednesday, August 19, 2015 5:30 PM
> To: Pledge Roy-R01356
> Cc: linuxppc-dev at lists.ozlabs.org; devicetree at vger.kernel.org; Bucur
> Madalin-Cristian-B32716; Wang Haiying-R54964
> Subject: Re: [PATCH 1/2] powerpc/qman: Change fsl,qman-channel-id to cell-
> index
> 
> On Wed, Aug 19, 2015 at 03:52:55PM -0500, Pledge Roy-R01356 wrote:
> > Sorry for digging up an old thread here Scott, but we never did close on this
> discussion.  See my replies inline below....
> >
> > > -----Original Message-----
> > > From: Wood Scott-B07421
> > > Sent: Tuesday, May 12, 2015 6:46 PM
> > > To: Pledge Roy-R01356
> > > Cc: linuxppc-dev at lists.ozlabs.org; devicetree at vger.kernel.org; Bucur
> > > Madalin-Cristian-B32716
> > > Subject: Re: [PATCH 1/2] powerpc/qman: Change fsl,qman-channel-id to
> > > cell- index
> > >
> > > On Tue, 2015-05-12 at 16:19 -0500, Pledge Roy-R01356 wrote:
> > > > > >
> > > > > > I don't believe this is correct - let me explain the rational
> > > > > > why we had two
> > > > > properties in the QMan portal to begin with.
> > > > > >
> > > > > > The two properties in question are cell-index and fsl,qman-channel-
> id.
> > > > > >
> > > > > > The cell-index property is used in u-boot as an index for the
> > > > > > software portal
> > > > > ID when adding the fsl,liodn from the U-boot table into the device
> tree.
> > > > >
> > > > > The device tree is not supposed to contain arbitrary software
> identifiers.
> > > >
> > > > I agree - this is why the original device tree bindings removed
> > > > cell-index as it can be calculated.
> > >  Unfortunately u-boot relied on
> > > > this value being present so to be backward compatible we don't
> > > > have a way to remove it.  I'm not sure on what the procedure is to
> > > > change things u-boot relies on,
> > >
> > > Generally the procedure is that we don't change it.  It wouldn't be
> > > so bad if using an old U-Boot just meant that datapath doesn't work
> > > with upstream kernels (since that doesn't work now), but a dts that
> > > makes existing U-Boot crash is another matter.
> >
> > If this is true then we can never remove the cell-index property.
> 
> Correct.
> 
> > The cell-index in this case is referring to the portal index which
> > could be calculated from the qman-portal at XXXX value.  My preference
> > would be to eliminate cell-index and replace it with this calculation
> > but that would mean older u-boot would fail to work with newer kernel.
> > While the bug that caused older u-boot to crash if this property is
> > annoying this has been addressed in more recent u-boots.  I can't
> > comment on a policy where u-boot must always boot newer version of
> > Linux - that means Linux will have to drag along baggage like this
> > property for a long time (forever?).
> 
> The baggage isn't particularly onerous.  It's just using a suboptimal property
> name.  The semantics are exactly the same as fsl,qman-channel-id.
> 
> > > > >
> > > > > > The  fsl,qman-channel-id property is used in Linux and
> > > > > > corresponds to a hardware value that indicates which channel
> > > > > > is dedicated to the software portal.
> > > > > >
> > > > > > While I'm not aware of a current SoC where the channel ID for
> > > > > > a software portal does not match the index (i.e. SWP 0 uses
> > > > > > channel 0,
> > > > > > etc.)
> > > > >
> > > > > Thus there's no backward compatibility issue with redefining
> > > > > cell-index to mean the channel ID.
> > > >
> > > > Channel IDs do change and are defined when the SoC is created
> > >
> > > But for SoCs that already exist, they won't change, right?  We don't
> > > need to care about what existing U-Boot does on new SoCs, since
> > > U-Boot would need to be changed to support the new SoC at all.
> >
> > This code isn't looking at SoC product numbers - the whole point of
> > putting this in the device tree is to avoid doing just that.  If we
> > had to add code for each SoC to u-boot we may as well get rid of the
> > device tree and hardcode this configuration in the source file that is
> > SoC specific.
> 
> The point of putting this in the device tree is to avoid per-SoC code in
> *Linux*.  U-Boot does use the device tree for similar purposes on some
> platforms, but that's not something we've done yet.  I'm not sure how it's
> relevant, though.  How would it be different if we had fsl,qman-channel-id
> and no cell-index?

I guess my point isn't getting through - channel-id and cell-index are too independent concepts that are coincidentally the same. Cell-index is only used by u-boot and is used to determine the portal number.  It is absolutely possible that we produce and SoC where portal number 0 has channel 0x1000.  We've changed these things in the past as well - this is why we introduced the channel-id property.  This is also consistent with other blocks (FMan, PME, DCE) that have channels associated with them.  

The cell-index is useless - the portal ID can be derived from the portal address.  The fact that you're unwilling to  remove cell-index doesn't mean that channel-id is redundant and should be removed.


> 
> When I mentioned U-Boot needing to be updated for new SoCs, I meant to
> run in general, not specifically the QMan code.  So, needing to be compatible
> with existing QMan code is only an issue for SoCs where an older U-Boot
> actually exists.  We don't need to care what the old code would have done
> on newer SoCs.
> 
> > > >  (look at checks for QMan versions and adjustments for Pool
> > > > Channel IDs in the driver).  If the channel ID for portal 0 ever
> > > > becomes non zero we just end up having to make a mess in the code
> > > > or reintroduce this field.
> > >
> > > What defines that portal as "portal 0"?
> >
> > Portal 0 is portal 0 because it is at offset 0 in the QMan portal
> > memory region.  Portal 1 is at 0x4000 etc...  Note that this is not
> > the case for forthcoming ARM devices as portals are distributed at 64K
> > intervals.  However since the device tree parsing code for ARM is
> > separate from the PPC code this will not pose any issue.
> 
> I don't understand why it would cause an issue in any case.  U-Boot would
> need to either know the mapping from portal address to channel id, or get it
> from the device tree.  There's no need to introduce the concept of "portal
> number" except perhaps as some internal implementation detail.
> 
> >
> > >
> > > > > > it is possible that future SoCs could stray from this model,
> > > > > > there is no reason for portal index to equal channel ID at all times.
> > > > >
> > > > > How can future SoCs dictate how we assign a software-defined
> > > identifier?
> > > > > If software wants it to be the same as the channel id, then it will be.
> > > > >
> > > > > If there is some aspect of the hardware itself (not the
> > > > > documentation) that cell-index currently corresponds to, other
> > > > > than the channel id, please make that clear.
> > > >
> > > > Channel ID is defined in the SoC RTL - it is not controlled by
> > > > software and it is not a software assigned identifier.  It is not
> > > > possible for SW to set these values.
> > >
> > > I said "other than the channel id".  In particular, I was asking
> > > about the concept of "portal index".
> >
> > The only thing cell-index indicates is the offset of the portal in the
> > QMan address space.
> 
> cell-index has been redefined to not mean that at all.  It now only means
> channel ID.  We can do this because the value happens to be the same for all
> existing SoCs (and we should be sure to avoid putting things into the SDK for
> the aforementioned ARM chips that are contrary to the new definition).
> 

U-boot doesn't dictate the HW architecture and shouldn't - you're trying very hard not to admit you changed something you didn't fully understand and actually making the system harder to deal with.  At no point in time have we ever assumed portal ID would equal channel id.  With your logic u-boot is broken because it is using channel ID to index into an array that is per portal not per channel.  We shouldn't have removed channel-id in the first place - trying to redefine that portals = channels is just plain incorrect.  If we can't remove cell-index then we have to live with that - trying to defend removing channel-id because for some cases portal-id = channel-id is wrong.  The fact that someone though cell-index was a good idea in the first place is the error - that data already exists in the device tree in the offset field.  We need channel-id to be consistent with other device tree nodes and because portal-id != cell-index.  I'm not sure how else to get through to you on this, your base assumption is wrong.

> -Scott