lkml.org 
[lkml]   [2015]   [Jan]   [26]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: N900 v3.19-rc5 arm atags_to_fdt.c is broken
On Mon, 26 Jan 2015, Pavel Machek wrote:

> Hi!
>
> > > > $ du -b arch/arm/boot/dts/omap3-n900.dtb
> > > > 70212 arch/arm/boot/dts/omap3-n900.dtb
> > > >
> > > > $ echo $((0x10000))
> > > > 65536
> > > >
> > > > I would say, problem is because omap3-n900 binary DT is too large
> >
> > I agree.
> >
> > > OK if that's the case, then your patch makes sense to me. It also
> > > seems we can have the temporary stack be larger than the initial
> > > stack just for atags_to_fdt.
> >
> > The stack size isn't the issue, but rather its location. We need to
> > position it away from the DT data. The DT size is known and we could
> > use that, plus some room for the insertion of new data coming from the
> > ATAG conversion.
> >
> > Something like the following would be a more robust solution:
> >
> > diff --git a/arch/arm/boot/compressed/head.S b/arch/arm/boot/compressed/head.S
> > index 68be901759..89718de9dd 100644
> > --- a/arch/arm/boot/compressed/head.S
> > +++ b/arch/arm/boot/compressed/head.S
> > @@ -263,16 +263,37 @@ restart: adr r0, LC0
> > * OK... Let's do some funky business here.
> > * If we do have a DTB appended to zImage, and we do have
> > * an ATAG list around, we want the later to be translated
> > - * and folded into the former here. To be on the safe side,
> > - * let's temporarily move the stack away into the malloc
> > - * area. No GOT fixup has occurred yet, but none of the
> > - * code we're about to call uses any global variable.
> > + * and folded into the former here. No GOT fixup has occurred
> > + * yet, but none of the code we're about to call uses any
> > + * global variable.
> > */
> > - add sp, sp, #0x10000
> > +
> > + /* Get the initial DTB size */
> > + ldr r5, [r6, #4]
> > +#ifndef __ARMEB__
> > + /* convert to little endian */
> > + eor r1, r5, r5, ror #16
> > + bic r1, r1, #0x00ff0000
> > + mov r5, r5, ror #8
> > + eor r5, r5, r1, lsr #8
> > +#endif
> > + /* 50% DTB growth should be good enough */
> > + add r5, r5, r5, lsr #1
> > + /* preserve 64-bit alignment */
> > + add r5, r5, #7
> > + bic r5, r5, #7
> > + /* clamp to 32KB min and 1MB max */
> > + movs r1, r5, lsr #15
> > + moveq r5, #(1 << 15)
> > + movs r1, r5, lsr #20
> > + movne r5, #(1 << 20)
>
> Dunno. Would it be easier/simpler to just use 1MB, always? Do we
> support machines with <16MB RAM?

If people are used to put other things relatively close to the kernel
image like, say, some initrd image, then I'd prefer to be more
conservative and avoid spreading out too much.


Nicolas


\
 
 \ /
  Last update: 2015-01-27 01:21    [W:0.190 / U:0.936 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site