randconfig bug: ARM/KVM link error in hyp_idmap section

[Arnd Bergmann]

On Thursday 29 January 2015 16:23:42 Christoffer Dall wrote:
> On Wed, Jan 28, 2015 at 8:57 PM, Arnd Bergmann <arnd at arndb.de> wrote:
>
> > 8<----
> > Subject: [PATCH] ARM: KVM: avoid "HYP init code too big" error
> >
> > When building large kernels, the linker will emit lots of veneers
> > into the .hyp.idmap.text section, which causes it to grow beyond
> > one page, and that triggers the build error.
> >
> 
> do you have a config handy that generates this error?

Attached to this mail. Use on linux-next

> > This moves the section into .rodata instead, which avoids the
> > veneers and is safe because the code is not executed directly
> > but always copied into a separate page first.
> >
> > I am unsure if I wrote this the correct way though, so
> > it needs to be reviewed carefully.
> >
> > Signed-off-by: Arnd Bergmann <arnd at arndb.de>
> >
> > diff --git a/arch/arm/kernel/vmlinux.lds.S b/arch/arm/kernel/vmlinux.lds.S
> > index ce01a2d3339f..f4de6e16d951 100644
> > --- a/arch/arm/kernel/vmlinux.lds.S
> > +++ b/arch/arm/kernel/vmlinux.lds.S
> > @@ -23,10 +23,14 @@
> >         VMLINUX_SYMBOL(__idmap_text_start) = .;                         \
> >         *(.idmap.text)                                                  \
> >         VMLINUX_SYMBOL(__idmap_text_end) = .;                           \
> > -       . = ALIGN(32);                                                  \
> > +       . = ALIGN(32);
> > +
> > +#define IDMAP_RODATA                                                   \
> > +       .rodata : {                                                     \
> >         VMLINUX_SYMBOL(__hyp_idmap_text_start) = .;                     \
> >         *(.hyp.idmap.text)                                              \
> > -       VMLINUX_SYMBOL(__hyp_idmap_text_end) = .;
> > +       VMLINUX_SYMBOL(__hyp_idmap_text_end) = .;                       \
> > +       }
> >
> >  #ifdef CONFIG_HOTPLUG_CPU
> >  #define ARM_CPU_DISCARD(x)
> > @@ -123,6 +127,7 @@ SECTIONS
> >         . = ALIGN(1<<SECTION_SHIFT);
> >  #endif
> >         RO_DATA(PAGE_SIZE)
> > +       IDMAP_RODATA
> >
> >         . = ALIGN(4);
> >         __ex_table : AT(ADDR(__ex_table) - LOAD_OFFSET) {
> >
> >
> the changes look ok, but I don't understand why putting stuff in rodata is
> a good solution, is it simply by chance that the linker then generates
> fewer veneers there?  I think we're only branching internally in the hyp
> idmap text page anyhow, so wondering why this appears in the first place...
> hmmm.

The linker will not generate any veneers for .rodata because it does not
expect executable code in there. As I said, above, this is also correct
because it matches how we access that section (read-only, never execute).

	Arnd
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 0x203B430A-config.gz
Type: application/gzip
Size: 26652 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/linux-arm-kernel/attachments/20150129/e2aa6fcd/attachment-0001.gz>