cvs commit: ports/devel/gold Makefile

[Ian Lance Taylor]

Juergen Lock <nox at jelal.kn-bremen.de> writes:

> Is this about `exec format error' (ENOEXEC)?  We discussed this
> yesterday on irc (hi Sprewell! :) and found out that gold can produce
> executables that won't run on FreeBSD 7 and 8 too (for him it was
> chromium and some test executables on 7.2/i386, for me it was test
> executables on 8/amd64, I couldnt reproduce it here on 8 and 7 i386) -
> and after some debugging using qemu's gdbstub and a FreeBSD 8/amd64
> livefs guest that I scp'd a problem executable into I found out that
> its this condition in sys/kern/imgact_elf.c that fails:
> 	http://fxr.watson.org/fxr/source/kern/imgact_elf.c#L721
>
>   719                 if (phdr[i].p_type == PT_INTERP) {
>   720                         /* Path to interpreter */
>   721                         if (phdr[i].p_filesz > MAXPATHLEN ||
>   722                             phdr[i].p_offset + phdr[i].p_filesz > PAGE_SIZE)
> 				  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
>   723                                 return (ENOEXEC);
>   724                         interp = imgp->image_header + phdr[i].p_offset;
>   725                         break;
>   726                 }
>   727         }
>
> (gdb) p/x phdr[1].p_offset 
> $11 = 0x1174
> (gdb) p/x phdr[1].p_filesz 
> $12 = 0x15
> (gdb) 
>
>  So it seems our kernel expects this data (path of the `image activator'
> aka shared linker aka /libexec/ld-elf.so.1) to be in the first page of
> the executable...

Technically an ELF executable is permitted to have the PT_INTERP
segment anywhere it likes, so I think the kernel is in error here.

However, it is good practice to put PT_INTERP in the first page of the
executable, and I have made that change to gold.  So this issue should
be fixed in the next release of the GNU binutils.

Ian