[U-Boot] [RFC] Booting Xen from a FIT - Additional discussion about a refactor

[Karl Apsite]

On 04/22/2015 09:55 PM, Simon Glass wrote:
> +Tom
> 
> Hi Karl,
> 
> On 22 April 2015 at 13:05, Karl Apsite <Karl.Apsite at dornerworks.com> wrote:
>> Hi!
>>
>> I work at DornerWorks with the Xen Hypervisor.  We work with a variety of
>> embedded systems, and we wanted to facilitate Xen's boot procedure through
>> U-boot's Flattened Image Tree (FIT) format.  I've already prototyped some of the
>> functionality we were hoping to see, so we thought it'd be prudent to begin a
>> discussion with denx to get your opinion on the matter,
>>
>> First Objective: (Summary of what was prototyped)
>> A Flattened Image Tree is capable of holding all of the necessary binaries
>> already, so we only need to make a quick change to allow u-boot to load an extra
>> binary (in this case, a linux kernel) so that Xen can boot and load the kernel
>> when it's ready.  I started by simply adding a line in the configuration of my
>> tree-source (.its) to look like:
>>
>> config at 1 {
>>     description = "Xen 4.6.0-unstable configuration";
>>     kernel = "xen_kernel at 1";
>>     fdt = "fdt at 1";
>>     gen_bin0 = "linux_kernel at 1";
>> };
>>
>> I investigated what effort would be needed to load the additional binary.
>>
>> Booting Xen is easy (only a kernel and fdt are required), but Xen will look at a
>> hard-coded memory address to try a load a linux kernel.  This has to be placed
>> in memory by u-boot.  The only major addition I needed, was to make u-boot care
>> about a config option named "generic-binary" and load it, no questions asked.  I
>> chose the name "generic binary" as I simply needed u-boot to load a [thing]
>> without any additional behavior.  I'm using it to specifically load a linux
>> kernel at a specific memory address in preparation for xen, but there could be
>> potential future uses, hence the ambiguous name.
> 
> I wonder whether you should add a new type for the target kernel?
> General binary seems a bit vague. Just a thought.

I do agree, I don't really like the term "generic binary" either.

When preparing to boot Xen, u-boot needs to take a binary, and simply put it in
place.  Unlike the other images/objects (kernel, fdt, ramdisk, etc) u-boot's
role is very simple in this regard: "Take these bits, and make sure they go over
here."

In this scenario, the action taken by u-boot should be agnostic to what the
image actually is.  U-boot should simply move a binary, without any additional
behavior.  This led me to choose a name just as generic.

Maybe changing the name to "Loadable(s)" would sound a bit better, but going so
far as to name it "kernel" could be misleading.

> 
>>
>> The hack is pretty simple, as most everything was in place to boot xen, and it
>> took a little extra legwork to make u-boot care about my gen_bin.  I added the
>> necessary structure to load another object using ramdisk functions as examples.
>>  By loading a separate binary, Xen was able to boot, and successively boot Linux
>> as expected.  If you have any thoughts on this process overall, we'd like to
>> take your concerns into consideration.
>>
>> For a little more fun, I extended out the configuration to be able to load N
>> number of generics.

This plan was part of the reason we were trying to keep the name more generic.
Keeping it generic allows for people other than xen users to take advantage of
the new feature in various ways.

>>
>> Affected Files:
>> common/bootm.c
>> common/image-fit.c
>> common/image.c
>> doc/uImage.FIT/source_file_format.txt
>> include/configs/xilinx_zynqmp.h
>> include/image.h
>>
>> Second Objective:
>> While I was there, I noticed that u-boot's binary loading logic isn't as modular
>> as I first expected.  Most objects eventually boil down to a boot_get_<thing>().
>> Example: boot_get_ramdisk() or boot_get_kernel().  These functions are all very
>> similar as they handle the various u-boot image types and load their specific
>> binary.  The functions appear to have grown similar in structure and operation
>> overtime.  I don't think it is too much to reduce the duplicated structure of
>> the functions into a common boot_get_object() function, while replacing some of
>> the extra function wrappings.
>>
>> Some quick diagrams to explain what I mean
>> Initial:  http://i.imgur.com/H44dXmN.png (29KB)
>> Refactor: http://i.imgur.com/apEpyWz.png (24KB)
> 
> SGTM.
> 
>>
>> The refactor will likely effect the following files, several of which contain
>> trivial name changes, or small changes in a variable type.
>> arch/arm/lib/bootm.c
>> arch/avr32/lib/bootm.c
>> arch/microblaze/lib/bootm.c
>> arch/mips/lib/bootm.c
>> arch/nds32/lib/bootm.c
>> arch/nios2/lib/bootm.c
>> arch/openrisc/lib/bootm.c
>> arch/powerpc/lib/bootm.c
>> arch/sh/lib/bootm.c
>> arch/sparc/lib/bootm.c
>> arch/x86/lib/bootm.c
>> common/bootm.c
>> common/bootm_os.c
>> common/cmd_bootm.c
>> common/image-fdt.c
>> common/image.c
>> include/bootm.h
>> include/image.h
>>
>> Summary
>> 1.) RFC: What is the opinion on implementing a FIT configuration option to
>> permit a FIT to boot Xen
> 
> Seems OK to me.
> 
>> 2.) RFC: What is the opinion on a potential refactor concerning the
>> image-loading functions (boot_get_ramdisk, boot_get_kernel, etc) in order to
>> unify them into one common boot_get_object() function.
> 
> That would be good. I spent a bit of time refactoring the code to
> reduce duplication - fit_image_load() was one of the results of that.
> If you think the code is duplicated now you should have seen it before
> :-)
> 
> Probably 'boot_get_image()' is better than 'boot_get_object()' given
> the current naming.
> 
> Please do add tests for the new functionality - see test/image for
> some existing tests. Python is preferred if the test is non-trivial.

Absolutely, I'll be sure to include some tests in the patch(es).

> 
> Regards,
> Simon
>