Re: [PATCH v6 2/4] x86/stackvalidate: Compile-time stack validation

[Josh Poimboeuf <jpoimboe@redhat.com>]

On Tue, Jul 07, 2015 at 04:35:17PM -0700, Andy Lutomirski wrote:
> On Tue, Jul 7, 2015 at 4:29 PM, Josh Poimboeuf <jpoimboe@redhat.com> wrote:
> > On Tue, Jul 07, 2015 at 03:57:14PM -0700, Andy Lutomirski wrote:
> >> On Tue, Jul 7, 2015 at 7:54 AM, Josh Poimboeuf <jpoimboe@redhat.com> wrote:
> >> >
> >> > It currently only supports x86_64.  I tried to make the code generic so
> >> > that support for other architectures can hopefully be plugged in
> >> > relatively easily.
> >> >
> >> > Currently with my Fedora config it's reporting over 1400 warnings, but
> >> > most of them are duplicates.  The warnings affect 37 .c files and 18 .S
> >> > files.  The C file warnings are generally due to inline assembly, which
> >> > doesn't seem to play nice with frame pointers.
> >>
> >> This issue might be worth bringing up on the gcc and binutils lists.
> >> If we need better toolchain support, let's ask for it.
> >
> > I think we found a good solution for this.  See my update at:
> >
> >   https://lkml.kernel.org/r/20150707223519.GA31294@treble.redhat.com
> 
> Does that force frame pointer generation?  If so, then once we have a
> real kernel unwinder, we might want a non-frame-pointer-forcing
> version for better code generation.  (That can wait, of course.)

I strongly doubt it would force frame pointer generation if
-fomit-frame-pointer is set.  But I'll verify :-)

> >> > +
> >> > +   This is a context switch instruction like sysenter or sysret.  Such
> >> > +   instructions aren't allowed in a callable function, and are most
> >> > +   likely part of kernel entry code.
> >> > +
> >> > +   If the instruction isn't actually in a callable function, change
> >> > +   ENDPROC to END.
> >> > +
> >> > +
> >> > +6. stackvalidate: asm_file.o: func()+0x26: jump to outside file from callable function
> >> > +   or
> >> > +   stackvalidate: asm_file.o: func()+0xd9: jump to dynamic address from callable function
> >> > +
> >> > +   These are constraints imposed by stackvalidate so that it can
> >> > +   properly analyze all jump targets.  Dynamic jump targets and jumps to
> >> > +   code in other object files aren't allowed.
> >>
> >> Does this not trigger due to optimized sibling calls to different files?
> >
> > This is a great point.  With CONFIG_FRAME_POINTER it's not a problem,
> > because it adds -fno-optimize-sibling-calls.
> >
> > Without it, I think stackvalidate would spit out a ton of "jump to
> > outside file" warnings.
> >
> > I haven't yet looked at the details of how exactly sibling calls work.
> > I'd assume they're disabled because they break frame pointers somehow.
> > Any idea if they'd also break DWARF CFI stack traces?
> 
> They'll certainly prevent unwinding from finding the pre-optimization
> caller, but the rest of unwinding should work.  I don't know why we
> turn it off, though.
> 
> You might want special-case jump-out-of-translation-unit to be okay if
> the stack frame is in its initial state.  That is:
> 
> func:
>      jmp elsewhere
> 
> could be considered okay, as could:
> 
> func:
>      push %rax
>      pop %rax
>      jmp elsewhere
> 
> and similar.

Ah, nice idea.  That might cover all the cases.  I'll try it.

> 
> >
> > I probably need to do some digging there.  If sibling calls don't break
> > CFI stack traces and we end up needing them, stackvalidate might need to
> > analyze the entire kernel image at once instead of its current per-.o
> > checking.
> >
> > Anyway, thanks a bunch for all your insightful feedback Andy!
> >
> 
> I'm just pretending to be insightful :)

Insightful or not, your comments have been very helpful!

-- 
Josh