[U-Boot] Default initrd and fdt load behaviour on ARM (Was: Re: [linux-sunxi] Re: Not able to boot ramdisk on cubietruck)

[Ian Campbell]

On Tue, 2014-03-11 at 13:17 -0400, Tom Rini wrote:
> On Tue, Mar 11, 2014 at 03:21:54PM +0000, Ian Campbell wrote:
> > On Tue, 2014-03-11 at 09:48 -0400, Tom Rini wrote:
> > > On Tue, Mar 11, 2014 at 10:38:05AM +0000, Ian Campbell wrote:
> > > > But what is the reason for the default behaviour of bootz doing things
> > > > which do not conform to linux/Documentation/arm/Booting and therefore is
> > > > not going to boot?
> > > 
> > > Because U-Boot is more than just ARM, and ARM boards not enforcing /
> > > getting the correct behavour via fdt_high / initrd_high or bootm_low /
> > > bootm_mapsize / bootm_size / CONFIG_SYS_BOOTMAPSZ.
> > 
> > So you consider this to be a board level problem, rather than an arch
> > level one? (IOW we can't just fix this for all arm boards at once)
> 
> I'm not 100% sure.  An issue is that lowmem isn't always 512MB, it may
> be 768MB.

I think in reality it can vary substantially between the two (or even
outside it) based on the kernel's .config (so outside u-boot's control).

>   That also may not really matter for this, I'm not sure.  But
> part of the problem is that AFAICS no one provides a 100% correct
> example today.  With some testing around now, for TI platforms (DDR
> starts at 0x80000000 it would be:
> /*
>  * We setup defaults based on constraints from the Linux kernel, which
>  * should also be safe elsewhere.  We have the default load at 32MB into
>  * DDR (for the kernel), FDT above 128MB (the maximum location for the
>  * end of the kernel), and the ramdisk 512KB above that (allowing for
>  * hopefully never seen large trees).  We say all of this must be within
>  * the first 512MB as that will always be within the kernel lowmem and
>  * thus visible via bootm_size.
>  */
> #define CONFIG_EXTRA_ENV_SETTINGS \
> -       "loadaddr=0x80200000\0" \
> -       "fdtaddr=0x80F80000\0" \
> -       "fdt_high=0xffffffff\0" \
> +       "loadaddr=0x82000000\0" \
> +       "fdtaddr=0x88000000\0" \
> +       "rdaddr=0x88080000\0" \
> +       "bootm_size=0x20000000\0" \
> 
> All of that said, I could be convinced that we could try and make
> common/image.c::getenv_bootm_size() smarter about defaults but we need
> to take a lot of care there, and allow people to opt-out as that's
> common code and how much lowmem we have is configurable a lot.

> > [...]
> > > > Should the global default be changed to either 0xffffffff (no
> > > > relocation) or to start-of-ram+256MB (which should satisfy the kernels
> > > > requirements without needing logic changes to handle "just after the
> > > > 128MB boundary" as a concept).
> > > > 
> > > > Or is it intended that each board should opt-in to a sensible default
> > > > via their default environment?
> > > > 
> > > > Does bootm suffer the same? I suspect it does, at least for fdt (since
> > > > initrd has a load addr in the u-boot header).
> > > 
> > > SoCs should set a sane and correct set of values, using either of the
> > > available mechanisms.  I think Tegra gets this all correct via
> > > CONFIG_SYS_BOOTMAPSZ and I've been meaning to whack the various TI parts
> > > as part of the generic distro work.
> > 
> > I took a look at BOOTMAPSZ on sunxi and it is using (16 << 20) which
> > seems sensible, if a bit small, so perhaps something else is going
> > wrong.
> > 
> > Tegra uses 256MB, might be worth giving that a go I guess.
> 
> Doing some playing around, BOOTMAPSZ isn't a complete solution, you need
> bootm_low also set.  In fact, I found the best results with just doing
> botom_low but there may be cases where you also want BOOTMAPSZ.

I'll put fiddling with this for sunxi on my todo list. Although I kind
of hope someone else will beat me to it ;-)

> > > And keep in mind that when we set fdt/initrd_high to a non-0xffffffff
> > > value we're setting a maximum that's still checked vs how much we have,
> > > so setting base+512MB is probably a best default for cases where we may
> > > run on boards with varying amounts of DDR.
> > 
> > I'm not sure I follow -- why is +512MB best?
> 
> I'm a fan of setting things to the maximum allowed, to allow for future
> growth.  This would also allow for pretty giant ramdisks to be passed
> which is unlikely but not impossible (I need to debug X and I can't use
> nfs/sd/etc for root, let me assemble a big ramdisk with everything I
> need).

That makes sense.

On the other hand the "just above 128MB" advice is even better since
that will Just Work and allow images to stretch to the end of RAM if
necessary. But I suppose that is a far bigger rework on the u-boot side
of things, to flip from "maximum address" allocation to "starting from"
allocation. I don't think I want to get into that...

>   Getting back to why this would be left up to the SoC/board
> level, if for some reason there's a reason to not go up quite that high,
> they can tweak it.

Ian.