Re: [PATCH v3] btrfs: should block unused block groups deletion work when allocating data space

[Wang Xiaoguang <wangxg.fnst@cn.fujitsu.com>]

hello,

On 09/09/2016 06:56 PM, Holger Hoffstätte wrote:
> On 09/09/16 12:18, Holger Hoffstätte wrote:
>> On Fri, 09 Sep 2016 16:17:48 +0800, Wang Xiaoguang wrote:
>>
>>> cleaner_kthread() may run at any time, in which it'll call btrfs_delete_unused_bgs()
>>> to delete unused block groups. Because this work is asynchronous, it may also result
>>> in false ENOSPC error.
>> <snip>
>>
>> With this v3 I can now no longer balance (tested only with metadata).
>> New chunks are allocated (as balance does) but nothing ever shrinks, until
>> after unmount/remount, when the cleaner eventually kicks in.
>>
>> This might be related to the recent patch by Naohiro Aota:
>> "btrfs: let btrfs_delete_unused_bgs() to clean relocated bgs"
>>
>> which by itself doesn't seem to do any harm (i.e. everything still seems
>> to work as expected).
> Actually even that is not true; both patches seem to be wrong in subtle
> ways. Naohiro's patch seems to prevent the deletion during balance, whereas
> yours prevents the cleaner from kicking in.
Indeed in my patch, I just change "struct mutex delete_unused_bgs_mutex"
to "struct rw_semaphore bg_delete_sem", and try to get bg_delete_sem when
we allocate data space, so this patch should work as before :)

>
> As a simple reproducer you can convert from -mdup to -msingle (to create
> bloat) and then balance with -musage=10. Depending on which of the two
> patches are applied, you end with bloat that only grows and never shrinks,
> or bloat that ends up in mixed state (dup and single).
Can you give me a simple test script to reproduce your problem.
(I can write it myself, but I'm afraid I may misunderstand you :) )

Regards,
Xiaoguang Wang
>
> Undoing both makes both balancing and cleaning work again.
>
> -h
>
>
>