Re: [Qemu-devel] out of bounds in set_cc_op()

[Paolo Bonzini]

On 21/12/2017 14:32, Laurent Vivier wrote:
> Le 21/12/2017 à 14:07, Laurent Vivier a écrit :
>> Le 21/12/2017 à 13:49, Thomas Huth a écrit :
>>> On 20.12.2017 22:56, Paolo Bonzini wrote:
>>>> On 20/12/2017 20:20, Peter Maydell wrote:
>>>>> On the x86/sanitizer build, new runtime errors:
>>>>>   GTESTER check-qtest-m68k
>>>>> /home/petmay01/linaro/qemu-for-merges/target/m68k/translate.c:230:12:
>>>>> runtime error: index -1 out of bounds for type 'const uint8_t [11]'
>>>>>
>>>>> ...and similar fails on one or two boards on most of the other
>>>>> guest architectures.
>>>>
>>>> These are preexisting bugs, now exposed by the boot-serial-test.
>>>> Thomas, can you identify the architectures that have a problem and
>>>> notify the maintainers?  In the meanwhile I'll keep the boot-serial-test
>>>> enhancements queued locally, and remove them from the pull request.
>>>
>>>  Laurent, Richard,
>>>
>>> looks like old_op is -1 when set_cc_op() is called here for the first
>>> time. The problem can be reproduced by running the mini-kernel directly.
>>> Just get http://people.redhat.com/~thuth/m68k-uart.bin and run QEMU like
>>> this:
>>>
>>>  qemu-system-m68k -nographic -kernel  ~/tmp/m68k-uart.bin -serial none
>>>
>>> That kernel only contains these few instructions:
>>>
>>>   0x41, 0xf9, 0xfc, 0x06, 0x00, 0x00,     /* lea 0xfc060000,%a0 */
>>>   0x10, 0x3c, 0x00, 0x54,                 /* move.b #'T',%d0 */
>>>   0x11, 0x7c, 0x00, 0x04, 0x00, 0x08,     /* move.b #4,8(%a0) */
>>>   0x11, 0x40, 0x00, 0x0c,                 /* move.b %d0,12(%a0) */
>>>   0x60, 0xfa                              /* bra.s  loop */
>>>
>>> The problem occurs during the second instruction (i.e. the first move.b).
>>>
>>> Do you have any ideas where this -1 in s->cc_op could come from?
>>
>> I think it comes from CCOp: it's the value of CC_OP_DYNAMIC.
>>
>> We should not use it to access cc_op_live[].
>>
>> I try to find a fix, but I think Richard knows this better than me.
> 
> This should fix the problem, but I'd like Richard checks it...
> 
> diff --git a/target/m68k/translate.c b/target/m68k/translate.c
> index b60909222c..721b5801da 100644
> --- a/target/m68k/translate.c
> +++ b/target/m68k/translate.c
> @@ -225,6 +225,11 @@ static void set_cc_op(DisasContext *s, CCOp op)
>      s->cc_op = op;
>      s->cc_op_synced = 0;
> 
> +    if (old_op == CC_OP_DYNAMIC) {
> +        tcg_gen_discard_i32(QREG_CC_OP);
> +        return;
> +    }

This tcg_gen_discard_i32 is correct, but all flags were potentially live
and can be discarded if the new op uses it(*).  So I'd replace "return"
with "old_op = CC_OP_FLAGS".

Paolo

        (*) in fact it's always true that all flags can be discarded.
            Only discarding some is an optimization to limit the number
            of generated ops.

>      /* Discard CC computation that will no longer be used.
>         Note that X and N are never dead.  */
>      dead = cc_op_live[old_op] & ~cc_op_live[op];
> 
> 
> Thanks,
> Laurent
>