629639 - elfhack: unsupported symbol in relocation when building with stack protection

Status: RESOLVED FIXED

(Firefox Build System :: General, defect)

Description

[Josh Matthews [:jdm]]

===
=== If you get failures below, please file a bug describing the error
=== and your environment (compiler and linker versions), and use
=== --disable-elf-hack until this is fixed.
===
/home/t_mattjo/src/firefox/mozilla-central/ff-dbg/build/unix/elfhack/elfhack -b test.so
test.so: terminate called after throwing an instance of 'std::runtime_error'
  what():  Unsupported symbol in relocation
make[6]: *** [test.so] Aborted (core dumped)
make[6]: *** Deleting file `test.so'

gcc version 4.4.4 20100630 (Red Hat 4.4.4-10) (GCC)
GNU ld version 2.19.51.0.14-37.fc12 20090722, coreutils 7.6-11
Fedora 12 x86

Comment 1

[Mike Hommey [:glandium]]

Could you give the output of readelf -r build/unix/elfhack/inject/*.o ?

Comment 2

[Vladimir LAVALLADE]

Looks related to my https://bugzilla.mozilla.org/show_bug.cgi?id=629638, posted my output of readelf there.

Comment 3

[Mike Hommey [:glandium]]

(In reply to comment #0)
> ===
> === If you get failures below, please file a bug describing the error
> === and your environment (compiler and linker versions), and use
> === --disable-elf-hack until this is fixed.
> ===
> /home/t_mattjo/src/firefox/mozilla-central/ff-dbg/build/unix/elfhack/elfhack -b
> test.so
> test.so: terminate called after throwing an instance of 'std::runtime_error'
>   what():  Unsupported symbol in relocation
> make[6]: *** [test.so] Aborted (core dumped)
> make[6]: *** Deleting file `test.so'
> 
> gcc version 4.4.4 20100630 (Red Hat 4.4.4-10) (GCC)
> GNU ld version 2.19.51.0.14-37.fc12 20090722, coreutils 7.6-11
> Fedora 12 x86

Could you check if "make check" fails the same way with http://hg.mozilla.org/users/mh_glandium.org/elfhack/ (which contains the same elfhack code) ? I can't reproduce with a live fedora 12 system with the same versions as those you wrote.

Comment 4

[Josh Matthews [:jdm]]

Looks like make check passes.

Comment 5

[Mike Hommey [:glandium]]

(In reply to comment #4)
> Looks like make check passes.

Could you check comment 1 ?

Comment 6

[Josh Matthews [:jdm]]

Attachment: readelf output

Comment 7

[Mike Hommey [:glandium]]

(In reply to comment #6)
> Created attachment 507960 [details]
> readelf output

00000014  00001403 R_386_GOT32       00000000   mcount

wtf is mcount ? Can you attach the output of objdump -d -r build/unix/elfhack/inject/x86.o ?

Comment 8

[Mike Hommey [:glandium]]

The preprocessed source would be helpful, too.

Comment 9

[Mike Hommey [:glandium]]

Forget it, I found what it is. http://lwn.net/Articles/264029/
I could trigger this with behaviour with the -pg option of gcc.

Comment 10

[Mike Hommey [:glandium]]

The patch I attached in bug 629638 would allow you to build with elfhack enabled.
This doesn't solve the root of the problem, though.

Comment 11

[Josh Matthews [:jdm]]

If the root of the problem is that I'm building with -pg, that would be my fault.  I forgot that I added that to my mozconfig to see if I could get gprof output from my build.

Comment 12

[Mike Hommey [:glandium]]

(In reply to comment #11)
> If the root of the problem is that I'm building with -pg, that would be my
> fault.  I forgot that I added that to my mozconfig to see if I could get gprof
> output from my build.

Well, it's your fault and it's not your fault. The injected code is in C because it's more portable than assembly, but it is not designed to be built with fancy stuff like lto, pg, etc.
So in any case, it should be built with very controlled flags.

Comment 13

[Jory A. Pratt [:anarchy]]

Attachment: readelf -r *.o

Mike I have tested patch from bug 629638 , and still see the same failure. If there is anything you would like tested let me know.

Comment 14

[Mike Hommey [:glandium]]

(In reply to comment #13)
> Created attachment 508740 [details]
> readelf -r *.o
> 
> Mike I have tested patch from bug 629638 , and still see the same failure. If
> there is anything you would like tested let me know.

000000000074  000c00000004 R_X86_64_PLT32    0000000000000000 __stack_chk_fail - 4
I guess this comes from one of the hardening build options, I guess stack protection. How are you passing these options ? Are they built-in defaults to the compiler ?

Comment 15

[Jory A. Pratt [:anarchy]]

(In reply to comment #14)
> (In reply to comment #13)
> > Created attachment 508740 [details]
> > readelf -r *.o
> > 
> > Mike I have tested patch from bug 629638 , and still see the same failure. If
> > there is anything you would like tested let me know.
> 
> 000000000074  000c00000004 R_X86_64_PLT32    0000000000000000 __stack_chk_fail
> - 4
> I guess this comes from one of the hardening build options, I guess stack
> protection. How are you passing these options ? Are they built-in defaults to
> the compiler ?
Mike, they are built into our hardened specs so the are automatically passed via the compiler.

Comment 16

[Jory A. Pratt [:anarchy]]

Attachment: ensure -fno-stack-protector is used for injection code

dummy.c
test.c
elf.cpp
elfhack.cpp
x86_64.c
x86_64-noinit.c
===
=== If you get failures below, please file a bug describing the error
=== and your environment (compiler and linker versions), and use
=== --disable-elf-hack until this is fixed.
===
test.so: Reduced by 12128 bytes
PASS

As you can see from the patch is a trivial work around, just need to ensure we pass -fno-stack-protector :)

Comment 17

[Mike Hommey [:glandium]]

Attachment: Build elfhack injected code without stack protection

This applies on top of bug 629638.

(In reply to comment #16)
> Created attachment 508875 [details] [diff] [review]
> ensure -fno-stack-protector is used for injection code
>
> As you can see from the patch is a trivial work around, just need to ensure we
> pass -fno-stack-protector :)

That was the patch from bug 629638 ;) but i get the idea

Comment 18

[Mike Hommey [:glandium]]

Sorry for the bug mess, it ended up being highjacked, but the original problem will be addressed in bug 629638.

Comment 19

[Mounir Lamouri (:mounir)]

Pushed:
http://hg.mozilla.org/mozilla-central/rev/8a376e2a5400