[U-Boot] [PATCH] ARM: rmobile: Convert to bootm_size

[Simon Goldschmidt <simon.k.r.goldschmidt@gmail.com>]

On Wed, Nov 28, 2018 at 1:19 PM Tom Rini <trini@konsulko.com> wrote:
>
> On Wed, Nov 28, 2018 at 09:46:43AM +0100, Simon Goldschmidt wrote:
> > On Wed, Nov 28, 2018 at 2:31 AM Tom Rini <trini@konsulko.com> wrote:
> > >
> > > On Tue, Nov 27, 2018 at 07:31:24PM +0100, Simon Goldschmidt wrote:
> > > > On Tue, Nov 27, 2018 at 4:47 PM Marek Vasut <marek.vasut@gmail.com> wrote:
> > > > >
> > > > > On 11/27/2018 04:26 PM, Simon Goldschmidt wrote:
> > > > > > On 27.11.2018 14:09, Marek Vasut wrote:
> > > > > >> On 11/27/2018 01:33 PM, Simon Goldschmidt wrote:
> > > > > >>> On Tue, Nov 27, 2018 at 1:25 PM Marek Vasut <marek.vasut@gmail.com>
> > > > > >>> wrote:
> > > > > >>>> On 11/27/2018 08:03 AM, Simon Goldschmidt wrote:
> > > > > >>>>> On Tue, Nov 27, 2018 at 1:11 AM Marek Vasut <marek.vasut@gmail.com>
> > > > > >>>>> wrote:
> > > > > >>>>>> Convert all Renesas R-Car boards to bootm_size of 256 MiB and drop
> > > > > >>>>>> both
> > > > > >>>>>> fdt_high and initrd_high. This change implies that the FDT and initrd
> > > > > >>>>>> will always be copied into the first 256 MiB of RAM instead of being
> > > > > >>>>>> used in place, which can cause various kinds of inobvious problems.
> > > > > >>>>>>
> > > > > >>>>>> The simpler problems include FDT or initrd being overwritten or being
> > > > > >>>>>> used from unaligned addresses, especially on ARM64. The overhead of
> > > > > >>>>>> copying the FDT to aligned location is negligible and these problems
> > > > > >>>>>> go away, so the benefit is significant.
> > > > > >>>>>>
> > > > > >>>>>> Regarding alignment problems with fitImage. The alignment of DT
> > > > > >>>>>> properties
> > > > > >>>>>> is always 32 bits, which implies that the alignment of the "data"
> > > > > >>>>>> property
> > > > > >>>>>> in fitImage is also 32 bits. The /incbin/ syntax plays no role
> > > > > >>>>>> here. The
> > > > > >>>>>> kernel expects all elements, including DT and initrd, to be
> > > > > >>>>>> aligned to
> > > > > >>>>>> 64 bits on ARM64, thus using them in place may not be possible.
> > > > > >>>>>> Using the
> > > > > >>>>>> bootm_size assures correct alignment, again with negligible overhead.
> > > > > >>>>> In my opinion, all of these raw addresses defined in scripts or config
> > > > > >>>>> should be removed: They are probably vulnerable to overwriting
> > > > > >>>>> themselves as they only provide an address, not a range.
> > > > > >>>> This is not an address, it's size. And this one at least assures that
> > > > > >>>> the first 256 MiB are reserved for the kernel/FDT/initrd during
> > > > > >>>> bootm time.
> > > > > >>> Sorry I did not express myself clear enough. I meant that "fdt_high"
> > > > > >>> and "initrd_high" are bad because they contain an address only, not a
> > > > > >>> range. The 'bootm_size' thing is much better!
> > > > > >> Well the fdt_high and intrd_high can also contain a special ~0 value,
> > > > > >> which says "use the fdt/initrd in place", which is dangerous.
> > > > > >>
> > > > > >>>>> Just out of curiosity: is it required to put fdt and initrd into the
> > > > > >>>>> first 256 MiB or is this just some 'random' limit to ensure we use lmb
> > > > > >>>>> but don't overwrite U-Boot (text, heap, stack, etc)? Because if so, my
> > > > > >>>>> series to fix the recent CVE issues improves lmb to not overwrite
> > > > > >>>>> U-Boot and other reserved addresses and you might be able to remove
> > > > > >>>>> 'bootm_size', too. The improved lmb code would just allocate an
> > > > > >>>>> aligned address somewhere in the available RAM.
> > > > > >>>> It's just the first 256 MiB from the beginning, so there's enough space
> > > > > >>>> between that and U-Boot on all these boards.
> > > > > >>> Of course. I wanted to know if it would be good enough if U-Boot would
> > > > > >>> just put it somewhere without overwriting things or do you really need
> > > > > >>> them in the first 256 MiB? Because the revised lmb code would make
> > > > > >>> sure there's nothing overwritten, so there would be no need to trim at
> > > > > >>> 256 MiB.
> > > > > >> You can put them anywhere, you just need to meet the alignment
> > > > > >> requirements. Can the new LMB code help somehow with that ? And if so,
> > > > > >> how ?
> > > > > >
> > > > > > My additions to the LMB code should only ensure nothing gets overwritten
> > > > > > so you don't have to limit boom_size to 256MiB (but use the complete RAM
> > > > > > when bootm_size is not set).
> > > > > > Alignment does not change but should already be OK with LMB as you use it?
> > > > >
> > > > > If I can use the entire RAM (except U-Boot and fitImage), that'd be
> > > > > nice. What change do I need to do ?
> > > >
> > > > I don't know yet, sorry. I basically asked this question to find out
> > > > about the usage of 'bootm_size'. It's not really documented and I
> > > > couldn't find out it's full meaning yet. Because e.g. it is set to 16
> > > > MiB for socfpga gen5, which sounds a little low...
> > > >
> > > > From reading the code, doesn't it already work when leaving out
> > > > 'bootm_size'? (And leaving out 'bootm_mapsize' as well and not
> > > > defining CONFIG_SYS_BOOTMAPSZ?)
> > >
> > > It's all a little funny, yes.  common/image.c is where all of this gets
> > > laid out and we keep everything being passed to bootm (whatever image
> > > type it may be) between bootm_low (either set explicitly or start of
> > > first bank of DRAM) and bootm_size (default ends up basically being all
> > > of DRAM) and then we do lmb stuff.
> > >
> > > The main reason, as I replied just now earlier in the thread, is that
> > > for Linux we need to make sure everything is in "lowmem" and rather than
> > > disabling relocation (which leads to people seeing fail to boots when
> > > something overlaps something else, over time, and they spend a bunch of
> > > time debugging) via initrd_high/fdt_high=0xffffffff SoCs should instead
> > > set a reasonable constraint on bootm_size so that we can make sure
> > > nothing overlaps, when we can do that.
> >
> > Right, I spent some time debugging this, too when starting with FIT
> > images containing Kernel, FDT, initrd and FPGA image...
> >
> > I wasn't aware of the requirement to have everything in "lowmem" for
> > Linux though (I'll even have to check what exactly the lowmem range is
> > for my platform/Kernel). And without this requirement, I just failed
> > to see why we need "boom_size" when it also works with defaulting to
> > just find a free memory block just somewhere in the DRAM.
>
> JFYI, the general lowmem limit is 768MB, but since it can be configured
> at build-time at least, it can actually get tiny (and off the top of my
> head, I don't think 256MB is as small as it can be, but rather a I think
> reasonable assumption on my part given the trade-offs in-kernel).

OK, thanks. I'll keep that in mind when testing my changes. And 256 MB
should probably enough for everything we need to to in U-Boot...

Simon