Re: [PATCH v2 2/8] kbuild: Support for Symbols.list creation

[Miroslav Benes <mbenes@suse.cz>]

> > Hmm, maybe I spoke too soon.
> >
> > I am having issues if I have a two-object livepatch module in which each
> > object file needs to specify its own KLP_MODULE_RELOC for the same
> > symbol name.
> >
> > For example: I have test_klp_convert.ko which is comprised of
> > test_klp_convert_a.o. which needs to refer to state_show,1 and
> > test_klp_convert_b.o. which needs to refer to state_show,2.
> >
> >   % make
> >   ...
> >     KLP     lib/livepatch/test_klp_convert.ko
> >   klp-convert: Conflicting KLP_SYMPOS definition: vmlinux.state_show,0 vs. vmlinux.state_show,1.
> >   klp-convert: Unable to load user-provided sympos
> >   make[2]: *** [scripts/Makefile.modpost:148: lib/livepatch/test_klp_convert.ko] Error 255
> >   make[1]: *** [/home/cloud-user/disk/linux/Makefile:1282: modules] Error 2
> >   make: *** [Makefile:170: sub-make] Error 2
> >
> > I take a closer look next week, but in the meantime, see the source
> > files below.
> 
> Okay, with a fresh set of eyes today, I think the issue can be
> summarized as:
> 
>   * "Special" livepatch symbols, as processed by klp-convert, require
>     external linkage, otherwise a new local storage instance will be
>     created and miss klp-convert altogether
> 
>   * When linking together two object files, their combined symbol table
>     will include a de-duped list of uniquly named global symbols
> 
> So if we are to run klp-convert on the combined module object file (as
> per v2 plus suggested changes in this thread), we are going to run into
> problems if ...
> 
> 
> Example
> =======
> 
> (quoted in previous reply), test_klp_convert_a and test_klp_convert_b
> have their own "state_show" variable in which each wishes to resolve to
> unique symbol positions (2 and 3 accordingly).
> 
> 
> test_klp_convert_a
> ------------------
> 
> We get one symbol table entry and one relocation as expected.
> 
>   % eu-readelf --symbols lib/livepatch/test_klp_convert_a.o
> 
>   Symbol table [27] '.symtab' contains 38 entries:
>    29 local symbols  String table: [28] '.strtab'
>     Num:            Value   Size Type    Bind   Vis          Ndx Name
>     ...
>      30: 0000000000000000      0 NOTYPE  GLOBAL DEFAULT    UNDEF state_show
>     ...
> 
>   % eu-readelf --relocs lib/livepatch/test_klp_convert_a.o
> 
>   Relocation section [12] '.rela.klp.module_relocs.vmlinux' for section [11] '.klp.module_relocs.vmlinux' at offset 0x4b8 contains 1 entry:
>     Offset              Type            Value               Addend Name
>     000000000000000000  X86_64_64       000000000000000000      +0 state_show
> 
> 
> test_klp_convert_b
> ------------------
> 
> Just like the other object file, one symbol table entry and one
> relocation:
> 
>   % eu-readelf --symbols lib/livepatch/test_klp_convert_a.o
>   Symbol table [24] '.symtab' contains 23 entries:
>    19 local symbols  String table: [25] '.strtab'
>     Num:            Value   Size Type    Bind   Vis          Ndx Name
>     ...
>      20: 0000000000000000      0 NOTYPE  GLOBAL DEFAULT    UNDEF state_show
>     ...
> 
>   % eu-readelf --relocs lib/livepatch/test_klp_convert_b.o
> 
>   Relocation section [ 9] '.rela.klp.module_relocs.vmlinux' for section [ 8] '.klp.module_relocs.vmlinux' at offset 0x118 contains 1 entry:
>     Offset              Type            Value               Addend Name
>     000000000000000000  X86_64_64       000000000000000000      +0 state_show
> 
> 
> test_klp_convert
> ----------------
> 
> But the combined test_klp_convert.klp.o file has only a single
> unresolved "state_show" symbol in its symbol table:
> 
>   % eu-readelf --symbols lib/livepatch/test_klp_convert.klp.o
> 
>   Symbol table [35] '.symtab' contains 57 entries:
>    47 local symbols  String table: [36] '.strtab'
>     Num:            Value   Size Type    Bind   Vis          Ndx Name
>     ...
>      52: 0000000000000000      0 NOTYPE  GLOBAL DEFAULT    UNDEF state_show
>     ...
> 
> though the .rela.klp.module_relocs.vmlinux relocation section has two
> entries:
> 
>   % eu-readelf --relocs lib/livepatch/test_klp_convert.klp.o
> 
>   Relocation section [17] '.rela.klp.module_relocs.vmlinux' for section [16] '.klp.module_relocs.vmlinux' at offset 0x446c8 contains 2 entries:
>     Offset              Type            Value               Addend Name
>     000000000000000000  X86_64_64       000000000000000000      +0 state_show
>     0x0000000000000010  X86_64_64       000000000000000000      +0 state_show
> 
> and it looks like the combined KLP_MODULE_RELOC still contains the two
> unique symbol position values (2 and 3):
> 
>   % objcopy -O binary --only-section=.klp.module_relocs.vmlinux lib/livepatch/test_klp_convert.klp.o >(hexdump -C)
>   00000000  00 00 00 00 00 00 00 00  02 00 00 00 00 00 00 00  |................|
>   00000010  00 00 00 00 00 00 00 00  03 00 00 00              |............|
>   0000001c

Nice :/
 
> Maybe we can work around this by modifying the annotation macros and/or
> klp-convert, or live with this for now.

The question is (and I'll check later. I cannot wrap my head around it 
now) if it at least works if there are two references of the same symbol 
in two different .o. It would be same state_show in this case and not two 
different ones. If it works then I think we can live with it for a while, 
because after all duplicate symbols are quite rare in the kernel.

I think it should not be a big problem to fix it though (famous last 
words). The information is there, so klp-convert with a changed annotation 
macro should deal with it.

> Note: I'm inclined to defer work on resolving this limitation to v4.  I
> would like v3 to focus on collecting and squashing the feedback up to
> now on v2.  There are a few other outstanding issues that I have run
> across while testing this patchset, so I feel that it would be clearer
> for folks to base comments on those off a clean v3 slate.

Makes sense to me.

Miroslav