Re: Global hotspare functionality

[Anand Jain <anand.jain@oracle.com>]

On 04/05/2016 03:32 AM, Yauhen Kharuzhy wrote:
> 2016-04-01 18:15 GMT-07:00 Anand Jain <anand.jain@oracle.com>:
>>>>> Issue 2.
>>>>> At start of autoreplacig drive by hotspare, kernel craches in
>>>>> transaction
>>>>> handling code (inside of btrfs_commit_transaction() called by
>>>>> autoreplace initiating
>>>>> routines). I 'fixed' this by removing of closing of bdev in
>>>>> btrfs_close_one_device_dont_free(), see
>>>>>
>>>>> https://bitbucket.org/jekhor/linux-btrfs/commits/dfa441c9ec7b3833f6a5e4d0b6f8c678faea29bb?at=master
>>>>> (oops text is attached also). Bdev is closed after replacing by
>>>>> btrfs_dev_replace_finishing(), so this is safe but doesn't seem
>>>>> to be right way.
>>>>
>>>>
>>>>    I have sent out V2. I don't see that issue with this,
>>>>    could you pls try ?
>>>
>>>
>>> Yes, it reproduced on v4.4.5 kernel. I will try with current
>>> 'for-linus-4.6' Chris' tree soon.
>>>
>>> To emulate a drive failure, I disconnect the drive in VirtualBox, so bdev
>>> can be freed by kernel after releasing of all references to it.
>>
>>
>>    So far the raid group profile would adapt to lower suitable
>>    group profile when device is missing/failed. This appears to
>>    be not happening with RAID56 OR there are stale IO which wasn't
>>    flushed out. Anyway to have this fixed I am moving the patch
>>     btrfs: introduce device dynamic state transition to offline or failed
>>    to the top in v3 for any potential changes.
>>    But firstly we need a reliable test case, or a very carefully
>>    crafted test case which can create this situation
>>
>>    Here below is the dm-error that I am using for testing, which
>>    apparently doesn't report this issue. Could you please try on V3. ?
>>    (pls note the device names are hard coded in the test script
>>    sorry about that) This would eventually be fstests script.
>
> Hi,
>
> I have reproduced this oops with attached script. I don't use any dm
> layer, but just detach drive at scsi layer as xfstests do (device
> management functions were copy-pasted from it).

  Nice. I was able reproduce this (also found lock dep issue when running
  this, since it was in the original code a separate patch was sent
  ou). The issue was due to that bdev wasn't null, to fix this the
  btrfs_device_enforce_state() is changed quite a bit. V4 is out.

Thanks, Anand