Re: linux-next: Tree for Mar 18 (objtool)

[Josh Poimboeuf]

On Wed, Mar 18, 2020 at 09:05:42PM +0100, Peter Zijlstra wrote:
> "Built-in Function: void __builtin_trap (void)
> 
>     This function causes the program to exit abnormally. GCC implements
>     this function by using a target-dependent mechanism (such as
>     intentionally executing an illegal instruction) or by calling abort.
>     The mechanism used may vary from release to release so you should
>     not rely on any particular implementation."
> 
> Sounds encouraging :-(
> 
> Clearly it now emits UD2, but who knows what it'll do next week.
> 
> I think what happens is that we have code that does:
> 
> 	if (cond)
> 		BUG()/WARN()
> 
> And ubsan figures that @cond is something undefined, and then we get
> transformed into:
> 
> 	if (cond) {
> 		__builtin_trap();
> 		BUG()/WARN()
> 	}
> 
> And presto, double UD2.
>
> Since x86 has GENERIC_BUG unconditionally, we should actually have
> __bug_table entries for each BUG()/WARN() instance, which should allow
> us to distinguish between a proper BUG()/WARN() and this
> __builtin_trap() nonsense.

... except the __builtin_trap() UD2 seems to be coming *after* the BUG
UD2.  Could it be the BUG UD2 itself which is convincing UBSAN to add
the __builtin_trap()?

If the BUG() macro were to use __builtin_trap() instead of the inline
asm UD2, GCC might be able to merge the two UD2's into one, and all
would be well.  But that could be tricky because of the __bug_table
stuff which needs to reference the UD2.

-- 
Josh