[PATCH 00/10] Updated ML300 & ML403 patches

[Peter Ryser]

> Hmm, did you use the ml403 and ml300 def configs?  What date did you 
> pull Linus' tree?  Kumar and Paul were talking today about some serial 
> subsystem breakage on the linux-2.6 tree this weekend... I'll fast 
> forward tonight and try it on my board. 

Okay, please let me know how this works for you.

> Try seeking to commit: 67daf5f11f06b9b15f8320de1d237ccc2e74fe43
> That's what I generated the latest patches against. 

Hmm, I only recently switched to using git. Is this number string some 
kind of a tag that I can synchronize my local git tree to? If so, how?

>> Anyway, there is another issue that I would like to bring up and it 
>> has to do with xparameters.h. The xparameters.h file, or more 
>> exactly, the xparameters_* file, is automatically generated by EDK 
>> and is then used to configure the devices in the Linux kernel at 
>> compile time. While I understand the desire to get away from a static 
>> device definition to device enumeration at run-time, the current set 
>> of patches is a step backwards for users from a useability point of 
>> view. Users will now have to modify xparameters*.h by hand which is 
>> an error-prone process. 
>
>
> Actually, users should *never* modifiy generated files.  The intent is 
> that board specific fixups go directly into the top level 
> xparameters.h so that newly generated files don't have to be touched.  
> But yes, I understand what you mean. 

An EDK user is free to choose arbitrary names for his peripherals. 
Additionally, Base System Builder uses different names for various 
boards (historically). With that it is impossible to make static 
assignments in xparameters.h. If you go back to the 2.4 kernel and have 
a look at xparameters_ml300.h you can see that the assignment of boards 
specific parameters to Linux specific parameters is done in there and 
that xparameters.h is basically used to chose the proper xparameters_* 
file for a given board.

>> Additionally, the original 'redefines' are now replaced with 
>> redefines in xparameters.h but differently for every board. I suggest 
>> we keep the 2.4 methodology until we can come up with a better 
>> approach to enumerate devices at run-time.
>
>
> Andrei & I are already discussing this.  I'm going to change the 
> xparameters redefines to provide a default set of mappings that can be 
> used if xparameters_*.h has the linux specific mappings. 

Thanks. Why not just use the xparameters_ml300.h file created by the 
system_linux.xmp in the EDK reference design for the ML403 and rename it 
to xparameters_ml403.h for inclusion into the kernel tree? We could then 
make a change in EDK, add a parameter that lets the user specify the 
board he uses, and with that automatically create an xparameters_ml403.h 
(or any other board for that matter).

> However, due to the fact that generated xparam files don't have the 
> Linux redefines if the FPGA engineer doesn't select a linux bsp.

That's not a recommended flow. It's very easy to create an EDK design 
with the proper settings and since it is very likely that things change 
during the design process of the FPGA the small investment into making 
the proper settings in the tool will save a lot of time in the end.

>   I think it's important to allow user defined 'fixups' for their 
> board. (I've personally worked on a couple of projects where the FPGA 
> engineer would not generate the Linux BSP).  Design specific fixups 
> can go into the top level xparameters.h without touching the generated 
> file 

I strongly believe that this approach fixes things in the wrong place. 
The correct thing to do is to use EDK to create a proper xparameters_*.h 
that matches the FPGA design. In your methodology, if the user decides 
to change the peripheral names in EDK he will have to go back and change 
the defines in xparameters.h. With the 2.4 kernel methodology that is 
not necessary as such changes will be represented in a regenerated 
board-specific xparameters_*.h

> <rant> BTW; it really bugs me that edk will generate different xparam 
> files depending on the bsp; why isn't there a single standard set of 
> data that is loaded into all xparam files; regardless of software 
> target?  Some no-OS targets need the same information that a Linux 
> port needs. </rant> 

EDK creates an xparameters.h that matches the names of the parameters in 
the hardware design. However, EDK is capable of assuming other 
personalities than 'standalone', for example Linux. With the Linux 
personality it creates the proper files AND directory structure for 
inclusion into the Linux kernel. Ideally, the source files that are used 
to create the Linux bsp for a given FPGA design should be included in 
the kernel tree and be maintained in there (maybe, in the xparameters 
directory). I'm not so sure though how well this would be accepted in 
the community. Opinions?

> I've avoided using the same names as used by the Linux redefines 
> because I don't know how stable the linux bsp naming convention is, 
> and I want to avoid a naming conflict.  If you can *guarantee* me that 
> those linux redefines are stable, then I have no problem using them 
> instead of the new defines that are currently in the patch.  If they 
> are not; then I'll just do a one-to-one mapping into a non-conflicting 
> namespace, and users can provide custom definitions as needed. 

The names are stable. They have not changed since xparameters_ml300.h 
has been initially published to the 2.4 repository and there are no 
intentions on changing them. And again, we really want to move towards a 
structure that allows for detecting peripherals at run-time. That will 
improve useability by a magnitude as no recompilation of the Linux 
kernel will be needed when the FPGA design changes.

> This really isn't a big deal anyway; most of this discussion will 
> become moot in short order.  Sometime in the next few releases, 
> linuxppc will flip over to using a flattened device tree to pass 
> device information from the boot loader to the kernel.  xparameters 
> will drop out of the kernel proper entirely except for the 
> edk-generated device drivers (which is another issue entirely).  All 
> the xparam stuff will be extracted into a device tree by u-boot or the 
> zImage wrapper.  The kernel just won't care.  :) 

I agree. That's the way to go. Let's work towards that goal and keep 
xparameters_* as they have been in 2.4 for the moment.

>> Specific to the patch: XPAR_DDR_SIZE is not the same as XPAR_MEM_*. 
>> XPAR_DDR_SIZE is specifically defined by the user as part of the BSP 
>> generation and indicates how much memory is available for Linux. This 
>> can be (and typically is) the same as the physically available memory 
>> but can be less than that. On the other hand XPAR_MEM_* can be the 
>> same or a multiple of the physically available memory (aliasing for 
>> cached and non-cached accesses). Statically defining the memory size 
>> in xparameters_ml403.h is not desirable. This is especially true for 
>> the multi-processor FPGA devices that might want to share the 
>> physically available memory between themselves.
>
>
> As you can see in embed_config.c; I already discovered this the hard 
> way   :( 

Right. Sorry, I was quoting the wrong file. The value should not be 
hard-coded in embed_config.c but instead XPAR_DDR_SIZE should be used 
which is defined in xparameters_ml403.h.

> Hmmm, I don't see any XPAR mem defines in xparameters_ml300.h.  (I 
> don't have a copy of the linux xparams for ml403 in front of me at the 
> moment)  Is this something new? 

I was referring to XPAR*MEM*, i.e. the base address and high address 
definition for the memory in EDK.

> Really, this isn't statically defined anyway.  The bootloader (u-boot 
> or zImage) passes the memory size into the kernel; and in fact the 
> kernel command line; or the board setup code can restrict the amount 
> of mem used by the kernel.  XPAR_MEM_* isn't used by the kernel proper 
> at all. 

Agreed.

> Thanks for the comments. 

Thanks for making this patch available. I know how much hard work it is 
to get this done.

>
>
> Another issue we need to discuss is if/how to support the xilinx 
> generated BSP in the kernel proper; but I'll leave that for a 
> different email. 

Okay.

> If there's enough interest; I'll setup another git tree for the virtex 
> specific patches. 

Hmm, interesting idea. Let's see what others think.

- Peter