On 04/18/2013 08:37 PM, Jens Axboe wrote:
> On Wed, Apr 17 2013, Steven Rostedt wrote:
>> On Wed, 2013-04-17 at 16:36 +0800, Wanlong Gao wrote:
>>> Hi Tj, all,
>>>
>>> I can get kernel panic on my machine with kernel 3.9.0-rc7-4-gbb33db7 every time
>>> after booting for several minutes.
>>>
>>> Here attached the panic picture and the config.
>>> I'm sorry that the panic log is not completed.
>>>
>>> But fortunately, the panic was gone when I reverted the below commit from Tj,
>>>
>>> commit 3a366e614d0837d9fc23f78cdb1a1186ebc3387f
>>> Author: Tejun Heo <tj@kernel.org>
>>> Date:   Fri Jan 11 13:06:33 2013 -0800
>>>
>>>     block: add missing block_bio_complete() tracepoint
>>>
>>> I think this will be helpful for you to resolve this bug, and this may be urgent,
>>> because it's already rc-7 now.
>>
>>
>> Doing a objdump on the code I found this:
>>
>> ffffffff8123d479:       8b 77 48                mov    0x48(%rdi),%esi
>> ffffffff8123d47c:       48 8d 7d c8             lea    -0x38(%rbp),%rdi
>> ffffffff8123d480:       48 8b 4d b8             mov    -0x48(%rbp),%rcx
>> ffffffff8123d484:       ba 28 00 00 00          mov    $0x28,%edx
>> ffffffff8123d489:       45 89 e8                mov    %r13d,%r8d
>> ffffffff8123d48c:       e8 1f d3 ea ff          callq  ffffffff810ea7b0 <trace_current_buffer_lock_reserve>
>> ffffffff8123d491:       48 85 c0                test   %rax,%rax
>> ffffffff8123d494:       49 89 c6                mov    %rax,%r14
>> ffffffff8123d497:       74 5a                   je     ffffffff8123d4f3 <ftrace_raw_event_block_bio_complete+0xb3>
>> ffffffff8123d499:       48 89 c7                mov    %rax,%rdi
>> ffffffff8123d49c:       e8 5f 4f ea ff          callq  ffffffff810e2400 <ring_buffer_event_data>
>> ffffffff8123d4a1:       48 8b 4b 10             mov    0x10(%rbx),%rcx
>> ffffffff8123d4a5:       31 d2                   xor    %edx,%edx
>> ffffffff8123d4a7:       48 85 c9                test   %rcx,%rcx
>> ffffffff8123d4aa:       74 02                   je     ffffffff8123d4ae <ftrace_raw_event_block_bio_complete+0x6e>
>>
>> ffffffff8123d4ac:       8b 11                   mov    (%rcx),%edx   <<<---- BUG IS HERE
>>
>> ffffffff8123d4ae:       89 50 08                mov    %edx,0x8(%rax)
>> ffffffff8123d4b1:       48 8b 13                mov    (%rbx),%rdx
>> ffffffff8123d4b4:       48 8d 78 20             lea    0x20(%rax),%rdi
>> ffffffff8123d4b8:       48 89 50 10             mov    %rdx,0x10(%rax)
>> ffffffff8123d4bc:       8b 53 30                mov    0x30(%rbx),%edx
>> ffffffff8123d4bf:       44 89 78 1c             mov    %r15d,0x1c(%rax)
>> ffffffff8123d4c3:       c1 ea 09                shr    $0x9,%edx
>> ffffffff8123d4c6:       89 50 18                mov    %edx,0x18(%rax)
>> ffffffff8123d4c9:       8b 53 30                mov    0x30(%rbx),%edx
>> ffffffff8123d4cc:       48 8b 73 20             mov    0x20(%rbx),%rsi
>>
>> According to the picture, that showed where in the code dump, it has the
>> command that died: "<8b> 11 89 50 08".
>>
>> Just before that, rcx was tested for zero, and passed (wasn't zero). But
>> it seems that it wasn't a real address space either.
>>
>> bio is allocated with kmalloc and not kzalloc, so it is possible that if
>> bio->bi_bdev never got initialized and is just garbage.
> 
> A bio is always fully initialized, regardless of which internal
> allocator it came from. If people are doing private kmallocs, then they
> better be using bio_init() as well.
> 
> Wanlong, would it be possible to get a full dmesg on boot see I can see
> what drivers and file systems are in use? Anything special about your
> setup.

Sure, attached.

Thanks,
Wanlong Gao

> 
> Will get this reverted for 3.9 final, so we have some time to
> investigate.
>