On 2/3/15 12:15 , Justin T. Gibbs wrote:
> My belief is that the race is eliminated from the normal dataset release 
> path.  What the previous patch doesn’t guard against is a dbuf_clear() 
> causing an immediate eviction of the dataset’s bonus buffer.  I need to trace 
> through the code further, but I believe this race is between a property 
> callback and dbuf_clear() being called when the dataset is destroyed.  Can 
> you confirm that the crashing systems are doing a lot of dataset destroy 
> operations?

The system in this case wasn't doing much in terms of explicit destroys.
It was doing dataset creations and in this case, we were renaming a
dataset. I'm not sure if renaming would cause us to enter the same paths
as destroy.

Robert

>> On Feb 3, 2015, at 12:49 PM, Robert Mustacchi <[email protected]> wrote:
>>
>> Hi,
>>
>> We've seen another variant of 5531 now pop up on our systems.
>> Specifically we're seeing dmu_buf_rele() get passed a NULL dmu_buf_t *.
>> It seems like we've only narrowed the race from 5531, but didn't manage
>> to kill it entirely.
>>
>> Specifically we're seeing the following stack trace:
>>
>>> $c
>> mutex_enter+0xb()
>> dmu_buf_rele+0x15(0, fffffffff7e00720)
>> dsl_dataset_rele+0x19(ffffff0340e250c0, fffffffff7e00720)
>> dsl_prop_notify_all_cb+0xbb(ffffff0254120900, ffffff02d13aa9c0, 0)
>> dmu_objset_find_dp+0x108(ffffff0254120900, 366, fffffffff7d742f0, 0, 2)
>> dsl_prop_notify_all+0x28(ffffff02654dbac0)
>> dsl_dir_rename_sync+0x2d1(ffffff00111a9b40, ffffff0260db3680)
>> dsl_sync_task_sync+0x10a(ffffff00111a9a60, ffffff0260db3680)
>> dsl_pool_sync+0x28b(ffffff0254120900, 1558)
>> spa_sync+0x27e(ffffff0254d3b000, 1558)
>> txg_sync_thread+0x227(ffffff0254120900)
>> thread_start+8()
>>
>>
>> The dump is available at
>> https://us-east.manta.joyent.com/Joyent_Dev/public/bugs/OS-3788/vmdump.0. I'd
>> appreciate if anyone had some time to look at this.
>>
>> Thanks,
>> Robert
>>
> 
_______________________________________________
developer mailing list
[email protected]
http://lists.open-zfs.org/mailman/listinfo/developer

Reply via email to