Re: [PATCH] tools: bpf: Disable stack protector

[Yonghong Song]

On 1/17/23 5:23 AM, Jose E. Marchesi wrote:
> > On 1/16/23 2:49 PM, Peter Foley wrote:
> > > On Mon, Jan 16, 2023 at 4:59 AM Eduard Zingerman <eddyz87@xxxxxxxxx> wrote:
> > > > A bit tangential, but since BPF LLVM backend does not support the
> > > > stack protector (should it?) there is also an option to adjust LLVM
> > > > to avoid this instrumentation, WDYT?
> > > That would probably be worth doing, yes.
> > > But given that won't help already released versions of clang, it
> > > should probably happen in addition to this patch.
> >
> > Peter,
> >
> > If I understand correctly (by inspecting clang code), the stack
> > protector is off by default. Do you have link to Gentoo build
> > page to show how they enable stack protector? cmake config or
> > a private patch?
> >
> > Jose,
> >
> > How gcc-bpf handle stack protector? The compiler just disables
> > stack protector for bpf target?
>
> It doesn't.  -fstack-protector is disabled by default in GCC.  When you
> use it you get something like:
>
>    $ echo 'int foo() { char s[256]; return s[3]; }' | bpf-unknown-none-gcc \
>      -fstack-protector -S -o foo.s -O2 -xc -
>    $ cat foo.s
>    	.file	"<stdin>"
>    	.text
>    	.align	3
>    	.global	foo
>    	.type	foo, @function
>    foo:
>    	lddw	%r1,__stack_chk_guard
>    	ldxdw	%r0,[%r1+0]
>    	stxdw	[%fp+-8],%r0
>    	ldxb	%r0,[%fp+-261]
>    	lsh	%r0,56
>    	arsh	%r0,56
>    	ldxdw	%r2,[%fp+-8]
>    	ldxdw	%r3,[%r1+0]
>   	jne	%r2,%r3,.L4
>    	exit
>    .L4:
>    	call	__stack_chk_fail
>    	.size	foo, .-foo
>    	.ident	"GCC: (GNU) 12.0.0 20211206 (experimental)"
>
> i.e. it pushes a stack canary and checks it upon function exit, calling
> __stack_chk_fail.
>
> If clang has -fstack-protector ON by default and you change the BPF
> backend in order to ignore the flag, I think we should do the same in
> GCC.

clang itself does not have -fstack-protector on by default. It is
hardened gentoo distribution unconditionally added -fstack-protector
to its clang distribution.

In clang/lib/Driver/ToolChains/Clang.cpp, we have
   ...
  // NVPTX doesn't support stack protectors; from the compiler's 
perspective, it
  // doesn't even have a stack!
  if (EffectiveTriple.isNVPTX())
    return;

and -fstack-protector is not effective for NVPTX. I guess we
could make it noop for BPF target as well.