Re: i386 kernel modules unusable due to .plt sections

From: Tijl Coosemans <tijl_at_FreeBSD.org>
Date: Fri, 27 Aug 2021 16:10:24 UTC
On Fri, 27 Aug 2021 11:32:43 -0400 Mark Johnston <markj@freebsd.org>
wrote:
> On Fri, Aug 27, 2021 at 05:29:34PM +0200, Tijl Coosemans wrote:
>> On Fri, 27 Aug 2021 17:24:58 +0300 Konstantin Belousov
>> <kostikbel@gmail.com> wrote:  
>>> On Fri, Aug 27, 2021 at 03:41:30PM +0200, Tijl Coosemans wrote:  
>>>> I use devel/llvm* to build base and just switched to llvm12.  It seems
>>>> that on i386 clang12 uses R_386_PLT32 relocations for some calls to at
>>>> least memset, memcpy and __stack_chk_fail (clang11 uses R_386_PC32).
>>>> These are converted to R_386_JMP_SLOT relocations by the linker which
>>>> aren't supported by the kernel, e.g. loading linux.ko gives "kldload:
>>>> unexpected relocation type" from sys/i386/i386/elf_machdep.c.  The PLT
>>>> entries also depend on a base pointer in %ebx but kernel modules aren't
>>>> compiled with -fPIC, so this can't work and I suspect this is a
>>>> regression in clang12.
>>>> 
>>>> The following code shows the difference between clang11 and clang12:
>>>> 
>>>> --------
>>>> #include <string.h>
>>>> 
>>>> void *
>>>> test_memset(void *p, int c, size_t len) {
>>>>         return (memset(p, c, len));
>>>> }
>>>> 
>>>> void *
>>>> test_memcpy(void *dst, const void *src, size_t len) {
>>>>         return (memcpy(dst, src, len));
>>>> }
>>>> 
>>>> void *
>>>> test_memmove(void *dst, const void *src, size_t len) {
>>>>         return (memmove(dst, src, len));
>>>> }
>>>> --------
>>>> 
>>>> Output of "readelf -r test.o" when compiled with "clang12 -c test.c -m32":
>>>> r_offset r_info   r_type              st_value st_name
>>>> 0000002c 00000504 R_386_PLT32         00000000 memset
>>>> 00000067 00000304 R_386_PLT32         00000000 memcpy
>>>> 000000a7 00000402 R_386_PC32          00000000 memmove
>>>> 
>>>> With clang11:
>>>> r_offset r_info   r_type              st_value st_name
>>>> 00000036 00000502 R_386_PC32          00000000 memset
>>>> 00000083 00000302 R_386_PC32          00000000 memcpy
>>>> 000000d2 00000402 R_386_PC32          00000000 memmove    
>>> 
>>> Are you asking (for somebody) to add R_386_JMP_SLOT to i386/elf_machdep.c?
>>> Like this, not even built.
>>> 
>>> diff --git a/sys/i386/i386/elf_machdep.c b/sys/i386/i386/elf_machdep.c
>>> index 3754b36d9e33..a26a4189e0ee 100644
>>> --- a/sys/i386/i386/elf_machdep.c
>>> +++ b/sys/i386/i386/elf_machdep.c
>>> @@ -245,6 +245,7 @@ elf_reloc_internal(linker_file_t lf, Elf_Addr relocbase, const void *data,
>>>  			break;
>>>  
>>>  		case R_386_GLOB_DAT:	/* S */
>>> +		case R_386_JMP_SLOT:
>>>  			error = lookup(lf, symidx, 1, &addr);
>>>  			if (error != 0)
>>>  				return (-1);  
>> 
>> No, I've tried that.  It handles the relocation but callers still don't
>> setup %ebx as PIC register.  I'm looking for someone to confirm it's a
>> compiler bug and not a missing flag or something.  I tried -fno-plt and
>> that has no effect.  
> 
> Does disabling -zifunc-noplt fix the problem?  I believe it's set by
> default for i386.

Removing it from sys/conf/kern.pre.mk didn't help.