Re: [PATCH 00/20] objtool: UACCESS validation v3

[Josh Poimboeuf]

On Thu, Mar 07, 2019 at 02:13:12PM +0100, Peter Zijlstra wrote:
> On Thu, Mar 07, 2019 at 01:55:26PM +0100, Peter Zijlstra wrote:
> > On Thu, Mar 07, 2019 at 01:03:17PM +0100, Peter Zijlstra wrote:
> 
> 
> > > 01be     20d3:  31 c0                   xor    %eax,%eax
> > > 01c0     20d5:  4c 39 eb                cmp    %r13,%rbx
> > > 01c3     20d8:  77 08                   ja     20e2 <__do_sys_waitid+0x1cd>
> 
> randconfig-build/kernel/exit.o: warning: objtool:   __do_sys_waitid()+0x1c3: (branch)
> 
> > > 01cd     20e2:	83 f0 01             	xor    $0x1,%eax
> > > 01d0     20e5:	48 89 c2             	mov    %rax,%rdx
> > > 01d3     20e8:	83 e2 01             	and    $0x1,%edx
> > > 01d6     20eb:	48 83 c2 02          	add    $0x2,%rdx
> > > 01da     20ef:	48 ff 04 d5 00 00 00 	incq   0x0(,%rdx,8)
> > > 01e1     20f6:	00 
> > > 01de 			20f3: R_X86_64_32S	_ftrace_branch+0x148
> > > 01e2     20f7:	84 c0                	test   %al,%al
> > > 01e4     20f9:	75 2d                	jne    2128 <__do_sys_waitid+0x213>
> > 
> > we do not take this branch and fall-through.
> 
> And that is the error, I think. We should've taken it and went to:
> 
>   return -EFAULT;
> 
> because:
> 
>  +1be	xor  %eax,%eax	eax=0
>  +1cd   xor  $0x1,%eax	eax=1
>  +1e2   test %al,%al	1&1==1 -> ZF=0
>  +1e4   jnz
> 
> Is an unconditional code sequence, but there's no way objtool can do
> that without becoming a full blown interpreter :/
> 
> > > 0213     2128:  49 c7 c7 f2 ff ff ff    mov    $0xfffffffffffffff2,%r15
> > > ffffffffffffe0eb }
> > > 021a     212f:  48 8d 65 d8             lea    -0x28(%rbp),%rsp
> > > 021e     2133:  4c 89 f8                mov    %r15,%rax
> > > 0221     2136:  5b                      pop    %rbx
> > > 0222     2137:  41 5c                   pop    %r12
> > > 0224     2139:  41 5d                   pop    %r13
> > > 0226     213b:  41 5e                   pop    %r14
> > > 0228     213d:  41 5f                   pop    %r15
> > > 022a     213f:  5d                      pop    %rbp
> > > 022b     2140:  c3                      retq

This "fixes" it, and also seems to help -Os make much code:

diff --git a/include/linux/compiler.h b/include/linux/compiler.h
index 445348facea9..8de63db58fdd 100644
--- a/include/linux/compiler.h
+++ b/include/linux/compiler.h
@@ -67,7 +67,7 @@ void ftrace_likely_update(struct ftrace_likely_data *f, int val,
 				.line = __LINE__,			\
 			};						\
 		______r = !!(cond);					\
-		______f.miss_hit[______r]++;					\
+		if (______r) ______f.miss_hit[1]++; else ______f.miss_hit[0]++; \
 		______r;						\
 	}))
 #endif /* CONFIG_PROFILE_ALL_BRANCHES */