Re: kldref: unhandled relocation type 2

On 18 Aug 2018, at 00:50, Steve Kargl <sgk_at_troutmask.apl.washington.edu> wrote:
> 
> On Fri, Aug 17, 2018 at 05:50:06PM -0400, Ed Maste wrote:
>> On 8 August 2018 at 13:27, Warner Losh <imp_at_bsdimp.com> wrote:
>>> 
>>>> % /usr/obj/usr/src/i386.i386/usr.sbin/kldxref/kldxref /boot/kernel
>>>> kldxref: unhandled relocation type 2
>>>> (40+ messages...)
>>> 
>>> 
>>> Oh, wait: relocation type, not module info.... That's not me, that's ed and
>>> the new linker I think, or Dimitry and the new clang...
>> 
>> 2 is R_386_PC32. It looks like the kernel's relocation code handles
>> this, but not kldxref. I hope to be able to look at it soon.
> 
> Thanks, Ed.  If you need any other information, just ask.

This rabbit hole is a bit deeper though.  Almost all .ko files in
/boot/kernel contain R_386_PC32 relocations:

$ for i in /boot/kernel/*.ko; do readelf -r $i | grep -q R_386_PC32 && echo $i; done | wc -l
     861

But kldxref *only* complains specifically about if_bge.ko, and it is not
clear to me why.

Maybe the R_386_PC32 relocations are first preprocessed away, somehow,
before they end up in kldxref's ef_reloc(), where the "unhandled
relocation type" warning is produced?

In any case, I tried the following patch, which at least makes the
warning go away, but might not be the right solution:

Index: usr.sbin/kldxref/ef_i386.c
===================================================================
--- usr.sbin/kldxref/ef_i386.c  (revision 337959)
+++ usr.sbin/kldxref/ef_i386.c  (working copy)
_at__at_ -82,11 +82,17 _at__at_ ef_reloc(struct elf_file *ef, const void *reldata,
                addr = (Elf_Addr)addend + relbase;
                *where = addr;
                break;
-       case R_386_32:  /* S + A - P */
+       case R_386_32:  /* S + A */
                addr = EF_SYMADDR(ef, symidx);
                addr += addend;
                *where = addr;
                break;
+       case R_386_PC32:        /* S + A - P */
+               addr = EF_SYMADDR(ef, symidx);
+               addr += addend;
+               addr -= (Elf_Addr)where;
+               *where = addr;
+               break;
        case R_386_GLOB_DAT:    /* S */
                addr = EF_SYMADDR(ef, symidx);
                *where = addr;
===================================================================

While here, I corrected the comment for R_386_32 relocations.  At least,
all the documentation I could find states that such relocations only add
the addend to the symbol, and the actual code also does that.

-Dimitry