Re: [PATCH][next] io_uring: ensure variable ret is initialized to zero

[Jens Axboe]

On 9/26/19 2:14 PM, Rasmus Villemoes wrote:
> On 26/09/2019 13.42, Jens Axboe wrote:
>> On 9/26/19 1:33 PM, Dan Carpenter wrote:
>>> On Thu, Sep 26, 2019 at 11:56:30AM +0200, Jens Axboe wrote:
>>>> On 9/26/19 11:50 AM, Colin King wrote:
>>>>> From: Colin Ian King <colin.king@xxxxxxxxxxxxx>
>>>>>
>>>>> In the case where sig is NULL the error variable ret is not initialized
>>>>> and may contain a garbage value on the final checks to see if ret is
>>>>> -ERESTARTSYS.  Best to initialize ret to zero before the do loop to
>>>>> ensure the ret does not accidentially contain -ERESTARTSYS before the
>>>>> loop.
>>>>
>>>> Oops, weird it didn't complain. I've folded in this fix, as that commit
>>>> isn't upstream yet. Thanks!
>>>
>>> There is a bug in GCC where at certain optimization levels, instead of
>>> complaining, it initializes it to zero.
>>
>> That's awfully nice of it ;-)
>>
>> Tried with -O0 and still didn't complain for me.
>>
>> $ gcc --version
>> gcc (Ubuntu 9.1.0-2ubuntu2~18.04) 9.1.0
>>
>> Tried gcc 5/6/7/8 as well. Might have to go look at what code it's
>> generating.
>>
> 
> I think it's essentially the same as
> https://lore.kernel.org/lkml/CAHk-=whP-9yPAWuJDwA6+rQ-9owuYZgmrMA9AqO3EGJVefe8vg@xxxxxxxxxxxxxx/
> (thread "tmpfs: fix uninitialized return value in shmem_link").

I think you're right, it's the same pattern. If I kill the:

if (ret)
	return ret;

inside the if (sig) branch, then gcc does show the warning as it should.

-- 
Jens Axboe