Re: [PATCH, RFC] xfs: fix multi-AG deadlock in xfs_bunmapi

[Nikolay Borisov <n.borisov.lkml@gmail.com>]

On 26.07.2017 16:04, Christoph Hellwig wrote:
> On Thu, Jul 20, 2017 at 05:58:42PM +0300, Nikolay Borisov wrote:
>>
>>
>> On 20.07.2017 10:49, Christoph Hellwig wrote:
>>> On Thu, Jul 20, 2017 at 10:47:50AM +0300, Nikolay Borisov wrote:
>>>>
>>>>
>>>> On 20.07.2017 10:47, Christoph Hellwig wrote:
>>>>> Odd.  And the patch to always just process a single extent from
>>>>> xfs_itruncate_extents fixes the issue for you, right?
>>>>
>>>> As far as I remember yes.
>>>
>>> Can you retest it under otherwise identical conditions?
>>
>> Okay, I confirm that indeed changing the define from 2 to 1 fixes the
>> issue.
> 
> But this is just on 4.4 and older, right?  We had issues in the
> allocation path there as well, which should be fixed in 4.11 and later,
> and in recrnt 4.9-stable releases.  Without those the allocation path
> might sometimes lock out of order as well.
> 

Couldn't repro on 4.11 and 4.10 (which supposedly shouldn't have the
fixes). I tried running generic/299 200 times on each kernel. Which
exactly are the patches in 4.11 which fix the alloc path?