Re: [GIT PULL] Core block IO bits for 2.6.39 - early Oops

[Jens Axboe <jaxboe@fusionio.com>]

On 2011-03-25 10:57, Markus Trippelsdorf wrote:
> On 2011.03.25 at 09:44 +0100, Jens Axboe wrote:
>> On 2011-03-25 09:37, Markus Trippelsdorf wrote:
>>> On 2011.03.25 at 08:23 +0100, Jens Axboe wrote:
>>>> On 2011-03-24 22:41, Markus Trippelsdorf wrote:
>>>>> On 2011.03.24 at 22:01 +0100, Jens Axboe wrote:
>>>>>> On 2011-03-24 21:06, Markus Trippelsdorf wrote:
>>>>>>> On 2011.03.24 at 20:57 +0100, Jens Axboe wrote:
>>>>>>>>
>>>>>>>> OK, still a data point. What was the last -git kernel you used?
>>>>>>>
>>>>>>> This one was the last and gave me no problems:
>>>>>>>
>>>>>>> commit b81a618dcd3ea99de292dbe624f41ca68f464376
>>>>>>> Merge: 2f284c8 a9712bc
>>>>>>> Author: Linus Torvalds <torvalds@linux-foundation.org>
>>>>>>> Date:   Wed Mar 23 20:51:42 2011 -0700
>>>>>>>
>>>>>>>     Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs-2.6
>>>>>>
>>>>>> Puzzling... Poking at straws here so far. Does this make any difference
>>>>>> whatsoever?
>>>>>
>>>>> I will test your patch later.
>>>>>
>>>>> Git-bisect gave me this result thus far:
>>>>>
>>>>> 9026e521c0da0731eb31f9f9022dd00cc3cd8885 is bad
>>>>> 82f04ab47e1d94d78503591a7460b2cad9601ede is good
>>>>>
>>>>> When I continue the bisection with 4345caba340f051e10847924fc078ae18ed6695c
>>>>> the system will start normally, but it then silently corrupts my xfs
>>>>> partitions. And on next (re)boot I get this (only fixable with
>>>>> xfs_repair):
>>>>>
>>>> How confident are you in those bisection results? Not trying to put you
>>>> on the spot, just wondering whether you tested and it's completely
>>>> consistent, or whether it was a one-off.
>>>
>>> Just double checked and 82f04ab47e1d94d78503591a7460b2cad9601ede is also
>>> bad. It just silently corrupts the file system (without a BUG) and I
>>> didn't notice.
>>> So back to square one. 
>>>
>>> How can I tell git-bisect just to try the commits in the block merge and
>>> not to take wild swings in history?
>>
>> Something like:
>>
>> $ git bisect start
>> $ git bisect good 3dab04e6978e358ad2307bca563fabd6c5d2c58b
>> $ git bisect bad 6c5103890057b1bb781b26b7aae38d33e4c517d8
> 
> Ok this time I've found the commit:
> 
> 9b6096a65f99a89dfd8328c4e469e7b53b3ae04a is the first bad commit
> commit 9b6096a65f99a89dfd8328c4e469e7b53b3ae04a
> Author: Shaohua Li <shaohua.li@intel.com>
> Date:   Thu Mar 17 10:47:06 2011 +0100
> 
>     mm: make generic_writepages() use plugging
>     
>     This recovers a performance regression caused by the removal
>     of the per-device plugging.
>     
>     Signed-off-by: Jens Axboe <jaxboe@fusionio.com>
> 
> Reverting it solves all problems here.

Can you try this one?

diff --git a/block/blk-core.c b/block/blk-core.c
index 59b5c00..8906ff1 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -1197,6 +1197,7 @@ static bool attempt_plug_merge(struct task_struct *tsk, struct request_queue *q,
 	if (!plug)
 		goto out;
 
+	preempt_disable();
 	list_for_each_entry_reverse(rq, &plug->list, queuelist) {
 		int el_ret;
 
@@ -1214,6 +1215,7 @@ static bool attempt_plug_merge(struct task_struct *tsk, struct request_queue *q,
 				break;
 		}
 	}
+	preempt_enable();
 out:
 	return ret;
 }

-- 
Jens Axboe