Hi Alan,
Thanks for taking the time to respond, I really appreciate your time.
I didn't quite follow all of the detail in your explanation, I'm
hoping you might be able to guide me through a little more as I really
do want to expand my understanding in this area.
Post by Alan ModraPost by Andrew BurgessPost by Alan ModraWhen --remove-relocations was added to objcopy with commit
d3e5f6c8f1e, objcopy lost the ability to remove dynamic relocation
sections such as .rela.plt from executables using the option
"--remove-section=.rela.plt". This patch reinstates that
functionality.
I thought it best to keep --remove-relocations as is, rather than
extending to handle dynamic relocations as per the patch in the PR,
because executables linked with --emit-relocs may have both dynamic
and non-dynamic relocations. In that case --remove-relocataions=* is
useful to remove all the non-dynamic relocations.
I think having an option that removes _almost_ all relocation sections
is going to cause user confusion. Yes, we can document it, but if we
can make it work for all relocation sections I think we probably
should.
Removing dynamic relocation sections from an executable is such a
weird thing to do that I'm not overly concerned. Also, they aren't
really removed. What actually happens is that the section disappears
from the section headers and the contents are overwritten with
zeros.
OK. That seems weird, but I feel like this is a deeper level of
knowledge, and I'm missing some of the simpler stuff, so, lets come
back to this later...
Post by Alan ModraContrast that with what happens to relocation sections in a
relocatable object, where the section is removed and other sections
take up its file space.
So, here you make a distinction between a 'relocatable object' and (I
assume) a 'non-relocatable object'.
So, I tried an experiment (this is all on x86-64 GNU/Linux):
$ cat test.c
#include <stdio.h>
volatile const char *message = "Hello World\n";
int
main ()
{
printf ("%s", message);
return 0;
}
$ gcc -o test.o -c test.c -g3 -O0
$ gcc --static -Wl,-emit-relocs -o test.x test.o -g3 -O0
$ readelf -WS test.x | grep rela.
[ 3] .rela.plt RELA 00000000004001d8 0001d8 0001f8 18 AI 0 38 8
[ 5] .rela.init RELA 0000000000000000 0b5d68 000018 18 I 57 4 8
[ 8] .rela.text RELA 0000000000000000 0b5d80 02c298 18 I 57 7 8
[10] .rela__libc_freeres_fn RELA 0000000000000000 0e2018 000f90 18 I 57 9 8
[12] .rela__libc_thread_freeres_fn RELA 0000000000000000 0e2fa8 000210 18 I 57 11 8
[15] .rela.rodata RELA 0000000000000000 0e31b8 011238 18 I 57 14 8
[17] .rela__libc_subfreeres RELA 0000000000000000 0f43f0 0000d8 18 I 57 16 8
[19] .rela__libc_IO_vtables RELA 0000000000000000 0f44c8 000fa8 18 I 57 18 8
[21] .rela__libc_atexit RELA 0000000000000000 0f5470 000018 18 I 57 20 8
[24] .rela__libc_thread_subfreeres RELA 0000000000000000 0f5488 000030 18 I 57 23 8
[26] .rela.eh_frame RELA 0000000000000000 0f54b8 005100 18 I 57 25 8
[29] .rela.tdata RELA 0000000000000000 0fa5b8 000060 18 I 57 28 8
[32] .rela.init_array RELA 0000000000000000 0fa618 000030 18 I 57 31 8
[34] .rela.fini_array RELA 0000000000000000 0fa648 000030 18 I 57 33 8
[36] .rela.data.rel.ro RELA 0000000000000000 0fa678 000270 18 I 57 35 8
[40] .rela.data RELA 0000000000000000 0fa8e8 000b40 18 I 57 39 8
[46] .rela.note.stapsdt RELA 0000000000000000 0fb428 0008d0 18 I 57 45 8
[48] .rela.debug_aranges RELA 0000000000000000 0fbcf8 000030 18 I 57 47 8
[50] .rela.debug_info RELA 0000000000000000 0fbd28 000630 18 I 57 49 8
[53] .rela.debug_line RELA 0000000000000000 0fc358 000018 18 I 57 52 8
[56] .rela.debug_macro RELA 0000000000000000 0fc370 0043e0 18 I 57 55 8
$ readelf -WS test2.x | grep rela.
[ 3] .rela.plt RELA 00000000004001d8 0001d8 0001f8 18 AI 0 37 8
[ 5] .rela.init RELA 0000000000000000 0b5d68 000018 18 I 56 4 8
[ 9] .rela__libc_freeres_fn RELA 0000000000000000 0b5d80 000f90 18 I 56 8 8
[11] .rela__libc_thread_freeres_fn RELA 0000000000000000 0b6d10 000210 18 I 56 10 8
[14] .rela.rodata RELA 0000000000000000 0b6f20 011238 18 I 56 13 8
[16] .rela__libc_subfreeres RELA 0000000000000000 0c8158 0000d8 18 I 56 15 8
[18] .rela__libc_IO_vtables RELA 0000000000000000 0c8230 000fa8 18 I 56 17 8
[20] .rela__libc_atexit RELA 0000000000000000 0c91d8 000018 18 I 56 19 8
[23] .rela__libc_thread_subfreeres RELA 0000000000000000 0c91f0 000030 18 I 56 22 8
[25] .rela.eh_frame RELA 0000000000000000 0c9220 005100 18 I 56 24 8
[28] .rela.tdata RELA 0000000000000000 0ce320 000060 18 I 56 27 8
[31] .rela.init_array RELA 0000000000000000 0ce380 000030 18 I 56 30 8
[33] .rela.fini_array RELA 0000000000000000 0ce3b0 000030 18 I 56 32 8
[35] .rela.data.rel.ro RELA 0000000000000000 0ce3e0 000270 18 I 56 34 8
[39] .rela.data RELA 0000000000000000 0ce650 000b40 18 I 56 38 8
[45] .rela.note.stapsdt RELA 0000000000000000 0cf190 0008d0 18 I 56 44 8
[47] .rela.debug_aranges RELA 0000000000000000 0cfa60 000030 18 I 56 46 8
[49] .rela.debug_info RELA 0000000000000000 0cfa90 000630 18 I 56 48 8
[52] .rela.debug_line RELA 0000000000000000 0d00c0 000018 18 I 56 51 8
[55] .rela.debug_macro RELA 0000000000000000 0d00d8 0043e0 18 I 56 54 8
This is a non-dynamic object (as far as I understand it) and notice
that relocation section .rela__libc_freeres_fn moved to offset 0xb5d80
to replace .rela.text once it was removed.
Using the same method I also tested building in these configurations:
$ gcc -fPIC -shared -Wl,-emit-relocs -o test.x test.c -g3 -O0
$ gcc -g3 -O0 -o test.x -Wl,-z,nocombreloc,--emit-relocs test.c
$ gcc -Wl,-emit-relocs -o test.x test.c -g3 -O0
In all cases, removing the .rela.text section caused another section
to move up to fill its place.
So, in my ignorance it appears (to me) that the distinction is between
dynamic relocs and non-dynamic relocs, not between relocatable and
non-relocatable objects.
Post by Alan ModraYou also might like to inspect the output of
gcc -o hello -Wl,-z,nocombreloc,--emit-relocs hello.c
before you discount the utility of the current --remove-relocations
behaviour.
As the person who added --remove-relocations I certainly have an
interest in its behaviour :)
As far as I can tell, --remove-relocations can remove anything except
dynamic relocations both before and after your patch.
It certainly was never my intention to introduce such an exception,
and if I had understood this limitation at the time of the original
patch I would have tried to remove the limitation.
However, as you can see above I included the 'nocombreloc' option in
one of my test cases, but without more of a clue I'm not sure what it
is that I'm supposed to be observing. As far as I can tell an object
built with that option responds to remove-relocations just like any
other.
Post by Alan ModraI call it a (perhaps accidental) feature rather than a bug
that the --emit-relocs created sections can be removed with
--remove-relocations.
I kind of feel like that's a pretty meta can-of-worms to open. If
emit-relocs didn't emit reloc sections that could be removed with
remove-relocations then I would have implemented remove-relocations
differently in order to allow that to be the case...
The original motivation behind --remove-relocations was that I had a
client who wanted to remove a relocation section. Before this option
the only relocation sections that could be removed were dynamic
relocation sections, something I didn't realise, and so I concluded
that NO relocation sections could be removed.
Then --remove-relocations was born and all was good. Except that I
managed to break the ability to remove dynamic relocation sections.
Post by Alan ModraPost by Andrew BurgessPost by Alan Modra/* Wrapper for dealing with --remove-section (-R) command line arguments.
A special case is detected here, if the user asks to remove a relocation
section (one starting with ".rela." or ".rel.") then this removal must
- be done using a different technique. */
+ be done using a different technique in a relocatable object. */
I agree that it's not 100% accurate, but it's near enough without
going into too much detail.
So, I started to write here:
".... I don't see what 'relocatable objects' have to do with this
at all, when it is 'dynamic relocations' that are the problem."
and then it struck me. I think what you're getting at is that,
without emit-relocs, the only relocations you'd expect to see in a
relocatable object are dynamic relocations, and so, in your comment,
relocatable object implies dynamic relocations.
However, wouldn't this mean that your comment should be reversed?
The original comment said that this function deals with requests to
remove a section, except that relocation sections must be handled
differently.
You've changed this to say that, .... relocation sections must be
handled differently in a relocatable object.
And if we apply the assumption that 'in a relocatable object' means
'for dynamic relocations', then this says that dynamic relocation
sections must be handled differently. Except this isn't true, dynamic
relocation sections must be handled just like any other section, it's
non-dynamic relocation sections that must be handled
differently. Right?
One further clarification, the client that I originally did the work
for uses --emit-relocs extensively, so when I wrote this code, and
when I've been thinking about it, I'm generally imagining an
object/executable with all the relocations preserved. This I guess,
is why I'm so keen to distinguish between different relocation types
requiring different handling, instead of different object types.
Thanks for your time,
Andrew