Re: Lock order reversal .

On Tue, Dec 7, 2010 at 9:18 AM, Julian Elischer <julian_at_freebsd.org> wrote:
> On 12/7/10 3:41 AM, Attilio Rao wrote:
>>
>> 2010/12/7 Erik Cederstrand<erik_at_cederstrand.dk>:
>>>
>>> Den 07/12/2010 kl. 10.20 skrev Garrett Cooper:
>>>
>>>> On Dec 7, 2010, at 12:26 AM, Mehmet Erol
>>>> Sanliturk<m.e.sanliturk_at_gmail.com>  wrote:
>>>>
>>>>> A Dmesg.TXT is attached having a lock order reversal .
>>>>
>>>>    The mount LOR is well known.
>>>
>>> I see that this is the standard response to lot's of LOR reports. It
>>> seems to be one of the most-reported errors on CURRENT (and it's certainly a
>>> loud one), but I think a lot of people waste time researching the error and
>>> browsing Bjoerns LOR page, only to get the above response (not picking on
>>> you, Garrett).
>>>
>>> Do we have the possibility of silencing well-known and presumably
>>> harmless LOR's if there isn't sufficient motivation to fix the source?
>>
>> Witness has an 'internal blessing list' we never wanted to use in
>> order to keep them popping up as reminder.
>> Actually, the fact the LOR is 'known' doesn't mean it is 'analyzed'.
>> The very few 'Analyzed but harmless' cases in the past have been
>> handled via _NOWITNESS flags I guess.
>
> the problem is that the witness output tells you the second case (the
> reversed case)
> but it doesn't have any clues about the first case (the one that wsa the
> other way around).
>
> An extended witness might use a lot of memory but associate with each lock a
> 'last place called when a lock was already held'
> that might give a clue as to where the other instance was. I'm not
> volunteering to write it,
> but it might be very worth while.. I'd certainly like to hear other ideas as
> well.

I have a small patch against stable/7 that adds a single bit to each
witness structure so that, if the "normal" lock order is ever
encountered after a reversal, the stack is printed.  It doesn't help
when the order is defined statically, though.

I could try to roll this up against -CURRENT this weekend.

Thanks,
matthew