On Wed, 16 May 2018, Alexey Dobriyan wrote:

> On Tue, May 15, 2018 at 09:25:53PM +0000, Anvin, H Peter wrote:
> > Why is that a problem?
> > Code: 00 00 00 eb e6 cc cc cc cc cc cc cc cc cc cc cc cc cc fa 8d b6 00 00 00 00 e8 5d e8 8f ff 8b 44 24 34 83 e0 03 83 f8 03 72 28 cc <cc> cc cc cc cc cc cc fa 8d b6 00 00 00 00 e8 3d e8 8f ff 89 e0
> > 
> > EIP: ret_from_intr+0xd/0x14 SS:ESP: 0068:8603bfb4
> > 
> > INT3 slipped through M586 => X86_ALIGNMENT_16 :-\
> 
> I could make the patch x86_64 only, but!
> 
> It crashed into the middle of the padding.
> 
> 796ef8fc <ret_from_intr>:
> 796ef8fc:       8b 44 24 34             mov    eax,DWORD PTR [esp+0x34]
> 796ef900:       83 e0 03                and    eax,0x3
> 796ef903:       83 f8 03                cmp    eax,0x3
> 796ef906:       72 28                   jb     796ef930 <resume_kernel>
> 796ef908:       cc                      int3
> 796ef909:       cc       <========>     int3

EIP points to the second int3 because that's where a simple return would
return to. do_trap() does not handle int3 special.

>From a quick check that seems to be the only code sequence which has:

     jcc     lbl

ENTRY(foo)

and ENTRY() aligns to a 16 byte boundary. So the simple fix for that is
below. We need objtool support to detect such places ...

Thanks,

	tglx

8<-------------------

--- a/arch/x86/entry/entry_32.S
+++ b/arch/x86/entry/entry_32.S
@@ -338,6 +338,11 @@ END(ret_from_fork)
 	cmpl	$USER_RPL, %eax
 	jb	resume_kernel			# not returning to v8086 or userspace
 
+	/*
+	 * Jump over the alignment padding which is filled with int3 instructions
+	 */
+	jmp	resume_userspace
+
 ENTRY(resume_userspace)
 	DISABLE_INTERRUPTS(CLBR_ANY)
 	TRACE_IRQS_OFF