[U-Boot] Unaligned flush_dcache_range in axs101.c

[Alexey Brodkin]

Hi Marek,

On Mon, 2016-04-11 at 19:54 +0200, Marek Vasut wrote:
> On 04/11/2016 07:48 PM, Alexey Brodkin wrote:
> > 
> > Hi Alex,
> > 
> > On Mon, 2016-04-04 at 09:38 +0200, Alexander Graf wrote:
> > > 
> > > Hi Alexey,
> > > 
> > > Marek just pointed out to me the fact that flush_dcache_range on arm
> > > expects cache line aligned arguments. However, it seems like in axs101.c
> > > we have an unaligned cache flush:
> > > 
> > >   flush_dcache_range(RESET_VECTOR_ADDR, RESET_VECTOR_ADDR + sizeof(int));
> > > 
> > > Could you please verify whether this is correct and if not just send a
> > > quick patch to fix it?
> > First this code is for support of Synopsys DesignWare AXS10x boards.
> > And AFAIK there's no such board that may sport ARM CPU instead or ARC.
> > So I'm wondering how did you bumped into that [issue?]?
> > 
> > Then I'm not really sure if there's a common requirement for arguments of
> > flush_dcache_range(). At least in "include/common.h" there's no comment about
> > that.
> Such comment should then be added. Sub-cacheline flush/invalidate calls
> can corrupt surrounding data.

Well this is not that simple really.

For example that's what we have on ARC:

[1] We may deal with each cache line separately. And BTW that's what we have
    now in U-boot, see http://git.denx.de/?p=u-boot.git;a=blob;f=arch/arc/lib/cache.c#l328
    In that case we only mention address of cache line start and regardless of its length
    line will be processed by HW.

[2] We may although deal with ranges as well (still this is not implemented in u-boot yet).
    In that case we need to set addresses of range beginning and end.
    But if start address falls actually in the middle of cache line it will be processed.
    And the same is valid for end of the region.

So from above I may conclude that it's more important to place data properly in memory.
I.e. if you put 2 completely independent substances in 1 cache line you won't be able to
deal with cache entries for them separately (at least on ARC).

I'm not really sure if ARM or any other arch in hardware may invalidate/writeback only part of
one cache line - that might very well be the case.

But in the original case my implementation makes perfect sense because what it does
it writes back instructions modified by the active CPU so others may see these changes.
and here I'd like ideally to have an option to write back only 1 CPU word (because that's
what I really modified) but this is not possible due to described above limitations of our HW.

-Alexey