From: Martin Hansson
Date: January 10 2011 3:59pm
Subject: Re: bzr commit into mysql-5.1-bugteam branch (martin.hansson:3520)
 Bug#58207
List-Archive: http://lists.mysql.com/commits/128328
Message-Id: <4D2B2CE2.9010200@oracle.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Content-Transfer-Encoding: 7bit

Jorgen Loland wrote:
>
>
> On 12/22/2010 03:49 PM, Martin Hansson wrote:
>> Jorgen Loland skrev 2010-12-16 12.27:
>>> On 12/16/2010 11:16 AM, Martin Hansson wrote:
>>>> Jorgen Loland skrev 2010-12-16 10.02:
>>>>>
>>>>>
>>>>> On 12/16/2010 09:21 AM, Martin Hansson wrote:
>>>>>> Jorgen Loland skrev 2010-12-15 14.07:
>>>>>>> Martin,
>>>>>>>
>>>>>>> The fix looks good, but the test case in the patch does not give
>>>>>>> valgrind warning. Please provide a test case that is fixed by the
>>>>>>> patch (e.g., the one reported in the bug). If you violently 
>>>>>>> refuse to
>>>>>>> replace this test, you can add one more instead.
>>>>>> You did build with BUILD/compile-pentium-valgrind-max, right? If you
>>>>>> still don't see it you might try to revert to the same revision as I
>>>>>> committed into.
>>>>>>
>>>>>> Does the test case in the bug report give valgrind warnings but mine
>>>>>> doesn't? If they do then that's serious enough, BUT... things 
>>>>>> actually
>>>>>> changed *while* I was working on the bug. First time around, the 
>>>>>> first
>>>>>> test caused a core dump. Then when I pulled the latest changes it
>>>>>> didn't, and then two days later it caused a crash again. The bug is
>>>>>> that
>>>>>> we get a totally runaway pointer, so it's sensitive to just about
>>>>>> anything.
>>>>>
>>>>> I compiled with compile-amd64-valgrind-max.
>>>>>
>>>>> Without your code changes, the test case in the bug report gives
>>>>> valgrind warnings like this:
>>>>>
>>>>> ==7710== Invalid read of size 1
>>>>> ==7710== at 0x4C28E61: memcpy (mc_replace_strmem.c:497)
>>>>> ==7710== by 0x76A3C3: create_tmp_table(THD*, TMP_TABLE_PARAM*,
>>>>> List<Item>&, st_order*, bool, bool, unsigned long long, unsigned long
>>>>> long, char*) (sql_select.cc:10412)
>>>>>
>>>>> These warnings go away when I apply your fix. The test case you added
>>>>> does not give valgrind warning before applying your code changes.
>>>>>
>>>> Revision number, please.
>>>
>>> 3524 gleb.shchepa@stripped
>>>
>> This is strange, I can repeat it just fine in the revision I committed,
>> 3524 and 3533. I do suspect you're running on 64 bit, though. Perhaps we
>> can work together to find a test case that fails on both architectures?
>> I currently don't have a 64 bit box set up. Only one is my mac on which
>> valgrind doesn't work.
>
> Yes, I'm running on 64 bit. The original test case seems to fail 
> "correctly" on my box.
>
Well that's good at least. So what can we do? Can I send you some tests 
and let you run them? I just spent an afternoon trying to get login to a 
decent 64 bit machine and for something this small, it doesn't seem 
worth it.

/Martin