Re: [PATCH bpf-next 0/2] bpf,ftrace: bpf dispatcher function fix

[Ilya Leoshkevich <iii@linux.ibm.com>]

On Tue, 2022-08-30 at 18:46 +0200, Ilya Leoshkevich wrote:
> On Tue, 2022-08-30 at 15:48 +0200, Jiri Olsa wrote:
> > On Tue, Aug 30, 2022 at 12:25:25AM +0200, Daniel Borkmann wrote:
> > > On 8/26/22 8:46 PM, Jiri Olsa wrote:
> > > > hi,
> > > > as discussed [1] sending fix that moves bpf dispatcher function
> > > > of out
> > > > ftrace locations together with Peter's
> > > > HAVE_DYNAMIC_FTRACE_NO_PATCHABLE
> > > > dependency change.
> > > 
> > > Looks like the series breaks s390x builds; BPF CI link:
> > > 
> > > https://github.com/kernel-patches/bpf/runs/8079411784?check_suite_focus=true
> > > 
> > >   [...]
> > >     CC      net/xfrm/xfrm_state.o
> > >     CC      net/packet/af_packet.o
> > >   {standard input}: Assembler messages:
> > >   {standard input}:16055: Error: bad expression
> > >   {standard input}:16056: Error: bad expression
> > >   {standard input}:16057: Error: bad expression
> > >   {standard input}:16058: Error: bad expression
> > >   {standard input}:16059: Error: bad expression
> > >     CC      drivers/s390/char/raw3270.o
> > >     CC      net/ipv6/ip6_output.o
> > >   [...]
> > >     CC      net/xfrm/xfrm_output.o
> > >     CC      net/ipv6/ip6_input.o
> > >   {standard input}:16055: Error: invalid operands (*ABS* and
> > > *UND*
> > > sections) for `%'
> > >   {standard input}:16056: Error: invalid operands (*ABS* and
> > > *UND*
> > > sections) for `%'
> > >   {standard input}:16057: Error: invalid operands (*ABS* and
> > > *UND*
> > > sections) for `%'
> > >   {standard input}:16058: Error: invalid operands (*ABS* and
> > > *UND*
> > > sections) for `%'
> > >   {standard input}:16059: Error: invalid operands (*ABS* and
> > > *UND*
> > > sections) for `%'
> > >   make[3]: *** [scripts/Makefile.build:249: net/core/filter.o]
> > > Error 1
> > >   make[2]: *** [scripts/Makefile.build:465: net/core] Error 2
> > >   make[2]: *** Waiting for unfinished jobs....
> > >     CC      net/ipv4/tcp_fastopen.o
> > >   [...]
> > >     CC      lib/percpu-refcount.o
> > >   make[1]: *** [Makefile:1855: net] Error 2
> > >     CC      lib/rhashtable.o
> > >   make[1]: *** Waiting for unfinished jobs....
> > >     CC      lib/base64.o
> > >   [...]
> > >     AR      lib/built-in.a
> > >     CC      kernel/kheaders.o
> > >     AR      kernel/built-in.a
> > >   make: *** [Makefile:353: __build_one_by_one] Error 2
> > >   Error: Process completed with exit code 2.
> > 
> > 
> > it does not break on my cross build with gcc 12, but I can
> > reproduce with gcc 8 (CI seems to be on gcc 9)
> > 
> > the problem seems to be wrong assembler code with extra '%'
> > that's generated for patchable_function_entry(5)
> > 
> > gcc 8 generates:
> > 
> > .LPFE1:
> >         nopr    %%r0
> >         nopr    %%r0
> >         nopr    %%r0
> >         nopr    %%r0
> >         nopr    %%r0
> > 
> > and gcc 12 generates:
> > 
> > .LPFE1:
> >         nopr    %r0
> >         nopr    %r0
> >         nopr    %r0
> >         nopr    %r0
> >         nopr    %r0
> > 
> > perhaps we need to upgrade gcc in CI? cc-ing Ilya, any idea?
> > 
> > thanks,
> > jirka
> 
> It's not obvious to me which gcc commit fixed this; I will bisect and
> find out. This will take some time.
> 
> However, officially, the kernel must be buildable by gcc 5.1+.
> Whatever I find, it's unlikely that we'll be able to backport it
> that far.
> 
> Therefore I think we need to find a way to conditionally
> do something else when using broken gccs. Or maybe just keep this
> x86-only after all.
> 
> Best regards,
> Ilya

FWIW, bisect points to

https://gcc.gnu.org/git/?p=gcc.git;a=commit;h=45d06a4045bebc3dbaaf0b1c676f4e22b7c6aca1

which makes perfect sense. Still, as I mentioned above, it's probably
worth tolerating brokens gccs instead of spending time backporting this
everywhere. And upgrading the CI machine will only paper over the
issue.

At a closer look, it looks weird to me that we have
patchable_function_entry(5) in a common header. If this optimization
is ever implemented for another architecture, a different number will
be required.

For simplicity, would it make sense to hide this under an #ifdef?
Something like this (untested):

#ifdef CONFIG_X86
#define BPF_DISPATCHER_ATTRIBUTES
__attribute__((patchable_function_entry(5)))
#else
#define BPF_DISPATCHER_ATTRIBUTES
#endif

Best regards,
Ilya