On Tue, Mar 24, 2020 at 11:44:33AM -0500, Josh Poimboeuf wrote:

On Fri, Mar 20, 2020 at 11:35:37AM -0700, Kees Cook wrote:

quoted

On Fri, Mar 20, 2020 at 09:26:13AM +0100, Peter Zijlstra wrote:

quoted

On Thu, Mar 19, 2020 at 12:45:50PM -0500, Josh Poimboeuf wrote:

quoted

quoted

On Thu, Mar 19, 2020 at 12:33:31PM -0500, Josh Poimboeuf wrote:

quoted

quoted

quoted

Actually I suspect it's the __builtin_unreachable() annotation which is
making UBSAN add the __builtin_trap()...  because I don't see any double
UD2s for WARNs.

quoted

Actually, removing __builtin_unreachable() *does* make the extra UD2 go
away -- I forgot I had some silly debug code.

LOL, check this:

"Built-in Function: void __builtin_unreachable (void)

    If control flow reaches the point of the __builtin_unreachable, the
    program is undefined. It is useful in situations where the compiler
    cannot deduce the unreachability of the code. "

Which, I bet, is what makes UBSAN insert that __builtin_trap().

What a friggin mess :/

What I'd like is to be able to specify to UBSAN what function to call
for the trap. I'd prefer to specify a well-defined exception handler,
but at present, UBSAN just inserts __builtin_trap().

Can't objtool be told to ignore a ud2 that lacks an execution path to it?

It can ignore unreachable UD2s, if we think that's the right fix.

I was hoping we could find a way to get rid of the double UD2s, but I
couldn't figure out a way to do that when I looked at it last week.

As far as I could tell, this needs patches to the UBSAN support in gcc
and clang. I have opened bugs for each:

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94307
https://bugs.llvm.org/show_bug.cgi?id=45295

-- 
Kees Cook