[Aboriginal] Sparc-32 doesn't work in 3.1.

Sergei Trofimovich slyich at gmail.com
Wed Dec 14 09:42:25 PST 2011


[ CCed Jakub ]

>>>   Boot time fixup v1.6. 4/Mar/98 Jakub Jelinek (jj at ultra.linux.cz).
>>>   Patching kernel for srmmu[Fujitsu TurboSparc]/iommu
>>>   Fixup i f029ddfc doesn't refer to a valid instruction at
>>>   f00de648[95eea000]
>>>   halt, power off

> I put the broken image up at http://landley.net/sparc-image for the
> moment, but if you build 3.1 with the attached .config and the toolchain
> mentioned last time, it should reproduce for you.  It's 100% reliable
> for me...

Nice! With this config it breaks for me on your and mine toolchains.
The offending function is ext4_kvmalloc (and similar ext4_kvzalloc).

The usual relocation in sparc looks like a pair of instructions loading
two pats of address in 2 instructions:

Like that:
>        sethi   %hi(ext4_fill_super), %o4       !, tmp113
>        or      %o4, %lo(ext4_fill_super), %o4  ! tmp113,, tmp28

In our case relocatable symbol sits in tail call, so %lo part is in "unusual"
RESTORE instruction:

> ext4_kvmalloc:
...
>         sethi  %hi(___i_page_kernel), %i2      !, tmp112
>        call    __vmalloc, 0    !
>         restore %i2, %lo(___i_page_kernel), %o2        ! tmp112,,
...

David: is this code correct? Or it's a compiler bug? I am sparc32 newbie.
(C source and asm sources of function are in [1])

I think this kind of code is generated only in -Os.
So to workaround it I tried this hack:

> --- a/fs/ext4/super.c
> +++ b/fs/ext4/super.c
>         ret = kmalloc(size, flags);
>         if (!ret)
>                 ret = __vmalloc(size, flags, PAGE_KERNEL);
> +
> +       asm __volatile__("nop":::"memory");
> +
>         return ret;
>  }
(for both ext4_kvmalloc / ext4_kvzalloc. Attached workaround as a patch.)

It forces compiler to geterate "usual" pattern for relocation.

I think of 2 solutions:

1. trying to fix sparc/boot/btfixupprep.c and arch/sparc/mm/btfixup.c
   to distinct HI22 and LO10 relocations as different ones.
   Right now they are merged into one 'i' type and rely on instruction heuristics to fix it.
2. Add a hack to arch/sparc/mm/btfixup.c to recognize restore instruction as well

Any others?

Hope that helps.

[1]:

void *ext4_kvmalloc(size_t size, gfp_t flags)
{
        void *ret;

        ret = kmalloc(size, flags);
        if (!ret)
                ret = __vmalloc(size, flags, PAGE_KERNEL);
        return ret;
}

        .global ext4_kvmalloc
        .type   ext4_kvmalloc, #function
        .proc   0120
ext4_kvmalloc:
        save    %sp, -96, %sp   !
        mov     %i0, %o0        ! size,
        call    __kmalloc, 0    !,
         mov    %i1, %o1        ! flags,
        cmp     %o0, 0  ! ret,
        bne     .LL274  !
         sethi  %hi(___i_page_kernel), %i2      !, tmp112
        call    __vmalloc, 0    !
         restore %i2, %lo(___i_page_kernel), %o2        ! tmp112,,
.LL274:
        jmp     %i7+8
         restore %g0, %o0, %o0  ! ret,
        .size   ext4_kvmalloc, .-ext4_kvmalloc

-- 

  Sergei
-------------- next part --------------
A non-text attachment was scrubbed...
Name: pessimizing-hack.patch
Type: text/x-patch
Size: 554 bytes
Desc: not available
URL: <http://lists.landley.net/pipermail/aboriginal-landley.net/attachments/20111214/7f6ae196/attachment-0003.bin>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 198 bytes
Desc: not available
URL: <http://lists.landley.net/pipermail/aboriginal-landley.net/attachments/20111214/7f6ae196/attachment-0003.pgp>


More information about the Aboriginal mailing list