Re: c6x linker issue on linux-next-20160808 + some linker table work

[Mark Salter]

On Tue, 2016-08-09 at 19:09 -0700, Luis R. Rodriguez wrote:
> On Aug 9, 2016 6:50 PM, "Mark Salter" <msalter@xxxxxxxxxx> wrote:
> >
> > On Tue, 2016-08-09 at 20:40 +0200, Luis R. Rodriguez wrote:
> > > On Tue, Aug 09, 2016 at 01:04:00PM -0400, Mark Salter wrote:
> > > >
> > > > On Tue, 2016-08-09 at 06:37 -0700, Guenter Roeck wrote:
> > > > >
> > > > > On 08/09/2016 01:11 AM, Luis R. Rodriguez wrote:
> > > > > >
> > > > > >
> > > > > > Mark, Aurelien,
> > > > > >
> > > > > > I've run into a linker (ld) issue caused by the linker table work I've
> > > > > > been working on [0]. I looked into this and for the life of me, I
> > > > > > cannot comprehend what the problem is, so was hoping you folks might
> > > > > > be able to chime in.
> > > > > >
> > > > > For reference, the error is
> > > > >
> > > > > c6x-elf-ld: drivers/built-in.o: SB-relative relocation but __c6xabi_DSBT_BASE not defined
> > > > > c6x-elf-ld: drivers/built-in.o: SB-relative relocation but __c6xabi_DSBT_BASE not defined
> > > > DSBT is a reference to the no-MMU userspace ABI used by c6x. The kernel shouldn't
> > > > be referencing DSBT base. The -mno-dsbt gcc flag should prevent it.
> > > I see -mno-dsbt on arch/c6x/Makefile already -- however at link time this is
> > > an issue if linker tables are used it seems. Do you have any other recommendation?
> > >
> > > I will note that it would seem that even i386 and x86-64 compiler/binutils seem
> > > to have relocation issues on older compiler/binutils, for instance:
> >
> > I see the problem with gcc 6 as well.
> >
> > So there appears to be some toolchain issues at play here. We build the kernel with two
> > c6x-specific options: -mno-dsbt and -msdata=none. I already mentioned dsbt. The sdata
> > option may be one of:
> >
> > -msdata=default
> > Â Â ÂPut small global and static data in the .neardata section, which is pointed to by
> > Â Â Âregister B14. Put small uninitialized global and static data in the .bss section,
> > Â Â Âwhich is adjacent to the .neardata section. Put small read-only data into theÂ
> > Â Â Â.rodata section. The corresponding sections used for large pieces of data are
> > Â Â Â.fardata, .far and .const.
> >
> > -msdata=all
> > Â Â Put all data, not just small objects, into the sections reserved for small data,
> > Â Â and use addressing relative to the B14 register to access them.
> >
> > -msdata=none
> > Â Â Make no use of the sections reserved for small data, and use absolute addresses
> > Â Â to access all data. Put all initialized global and static data in the .fardata
> > Â Â section, and all uninitialized data in the .far section. Put all constant data
> > Â Â into the .const section.
> >
> >
> > Both small data and DSBT make use of base register + 15-bit offset to access data
> > and thus the SB-relative reloc in the above error message.
> >
> > I think that gcc sees the .rodata section from DEFINE_LINKTABLE_RO() for builtin_fw
> > and thinks it needs an SB-relative reloc. When the linker sees that reloc, it thinks
> > it needs the dsbt base register and thus the error. Interestingly, weak data is
> > never put in the small data section so if gcc sees that data is weak, it doesn't
> > check the section name to see if it is a small data section. So SB-relative only
> > gets used for builtin_fw__end, but not the weak builtin_fw even though they both
> > are in the .rodata section.
> >
> > I suspect gcc should avoid being fooled by .rodata if -msdata=none is used.
> > Regardless, I think this could all be avoided if the RO tables used .const
> > instead of .rodata for c6x.
> Thanks for the thorough analysis, would you be OK for c6x to use .const for all read only linker tables or section ranges ?
> I had not added #ifndef around the core-sections.h main ELF definitons but could add one as its needed. In this case perhals that is needed and fine by you
> for SECTION_RODATA.
> We can also override any of the core section setter helpers for archs but in this case based on what you say it seems this is needed. Unless of course just
> -msdata=none is fine and that's not yet used and you prefer that.
> Â Luis

We're already using -msdata=none for kernel builds. From the gcc docs, one would think
all const data goes into .const with -msdata=none, but the kernel forces a lot of weak
const kallsyms data ,rodata so c6x vmlinux.lds still needs to have a .rodata section. I
think we need to use .const for the c6x read-only linker tables and keep .rodata for
RO_DATA_SECTION in vmlinux.lds.h.