[patch] Wine DLL base address patches

[Konstantin Belousov]

On Tue, Feb 26, 2013 at 10:52:15PM +0200, Damjan Jovanovic wrote:
> On Fri, Feb 22, 2013 at 5:19 AM, Damjan Jovanovic <damjan.jov at gmail.com> wrote:
> > On Thu, Feb 21, 2013 at 5:44 PM, Konstantin Belousov
> > <kostikbel at gmail.com> wrote:
> >> On Thu, Feb 21, 2013 at 12:57:45AM +0200, Damjan Jovanovic wrote:
> >>> On Wed, Feb 20, 2013 at 10:51 PM, Tijl Coosemans <tijl at coosemans.org> wrote:
> >>> > On 20-02-2013 16:48, Konstantin Belousov wrote:
> >>> >> On Wed, Feb 20, 2013 at 05:29:01PM +0200, Damjan Jovanovic wrote:
> >>> >>> Hi
> >>> >>>
> >>> >>> Wine needs some of its libraries to be loaded at specific base
> >>> >>> addresses (https://wiki.freebsd.org/Wine), something FreeBSD currently
> >>> >>> lacks.
> >>> >>>
> >>> >>> I've written a patch to the dynamic loader (/libexec/ld-elf.so.1) that
> >>> >>> loads libraries at their preferred base addresses
> >>> >>> (http://www.freebsd.org/cgi/query-pr.cgi?pr=176216), as well as a port
> >>> >>> of Prelink to FreeBSD which Wine uses to set base addresses
> >>> >>> (http://www.freebsd.org/cgi/query-pr.cgi?pr=176283). Both work :-),
> >>> >>> the changed dynamic loader doesn't show any problems in a few days of
> >>> >>> testing, and prelink works with the --reloc-only option as used by
> >>> >>> Wine.
> >>> >>>
> >>> >>> Please review/test/comment/commit.
> >>> >>
> >>> >> Unfortunately, it is not safe. MAP_FIXED overrides any previous mappings
> >>> >> which could exist at the specified address.
> >>> >
> >>> > I've simplified the rtld patch to a single line. The second patch makes
> >>> > Wine use -Ttext-segment linker flag instead of prelink. This requires
> >>> > binutils from ports, but it's easier than porting prelink.
> >>> >
> >>>
> >>> All of that occurred to me as well.
> >>>
> >>> The problem with that one-line rtld patch is that loading an
> >>> application will now fail if any of its libraries cannot be loaded at
> >>> their requested address.
> >> But this is intended behaviour. Also, the default virtaddr base for the
> >> shared libraries is 0, so the existing binaries should be not affected.
> >
> > In that case, and since failing to load a library only causes the
> > process to exit when starting up and not when it calls dlopen(), I
> > approve.
> >
> >>>
> >>> The problem with -Ttext-segment (and isn't it just -Ttext?) is that it
> >>> doesn't seem to work: the base_vaddr seen by rtld will remain 0, and
> >>> the address listed in /proc/.../map is different from what it should
> >>> be. Also run "readelf -l" on a library compiled that way and compare
> >>> with the output of one run through "prelink --reloc-only", you'll see
> >>> the lowest VirtAddr and PhysAddr in LOAD headers change only with
> >>> prelink. I really ported prelink because there was no other choice.
> >> The -Ttext-segment does work. As indicated by Tijl, you need recent
> >> binutils.  I just verified that ld 2.32.1 obeys -Ttext-segment.
> >>
> >> You can also take a look at the default linker script to see how
> >> -Ttext-segment is used, look for SEGMENT_START("text-segment").
> >>
> >
> > My apologies: I confused -Ttext which is documented but doesn't work,
> > with -Ttext-segment which is undocumented in FreeBSD 9.1 and might
> > work. I would test it further, but -CURRENT doesn't installworld
> > (ERROR: Required auditdistd user is missing, see /usr/src/UPDATING.)
> > and I am away until next week.
> >
> > Prelink is now in Ports. What I'd recommend is checking if the
> > binaries are the same, and if not, doing a diff between "readelf -a"
> > outputs of the prelinked binary vs -Ttext-segmented binary. Also run
> > this a few times and make sure the address is what's expected:
> >
> > #include <windows.h>
> > #include <stdio.h>
> > int main(int argc, char **argv)
> > {
> >   printf("%p\n", LoadLibrary("KERNEL32"));
> >   return 0;
> > }
> >
> > mingw32-gcc hello.c -o hello.exe
> > wine hello.exe
> 
> 
> With binutils 2.23.1 (in ports), comparing the output of "ld
> -Ttext-segment=0x7b800000" and "prelink --reloc-only 0x7b800000" using
> diffs of "readelf -a" outputs gives this:
> -    11: 000000007b800000     0 OBJECT  LOCAL  DEFAULT    6
> _GLOBAL_OFFSET_TABLE_
> +    11: 0000000000000000     0 OBJECT  LOCAL  DEFAULT    6
> _GLOBAL_OFFSET_TABLE_
> in other words, prelink also shifts the global offset table to the
> requested base address, ld does not. I don't think this matters since
> it's only ELF segments that get loaded - sections are irrelevant.
I suspect that it is sort of bug in ld. On the other hand, _G_O_T_ symbol
should be not used for real relocations, because both i386 and amd64 define
specific relocations which allow to reference the begining of the GOT.
The symbol is mostly a symbolic way to generate the relocations.

So indeed, this should be fine.
> 
> "objdump -s" finds no differences.
> 
> So I am happy with all of Tijl's patches, please commit them.

I expect Tijl to do it himself.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 834 bytes
Desc: not available
URL: <http://lists.freebsd.org/pipermail/freebsd-hackers/attachments/20130227/b0914cb7/attachment.sig>