Re: [PATCH 2/5] Btrfs: fix trans block rsv regression

[Liu Bo <bo.li.liu@oracle.com>]

On 09/14/2012 09:01 PM, Liu Bo wrote:
> On 09/14/2012 08:41 PM, Josef Bacik wrote:
>> On Fri, Sep 14, 2012 at 02:58:04AM -0600, Liu Bo wrote:
>>> In some workloads we have nested joining transaction operations,
>>> eg.
>>>   run_delalloc_nocow
>>>       btrfs_join_transaction
>>>       cow_file_range
>>>          btrfs_join_transaction
>>>
>>> it can be a serious bug since each trans handler has only two
>>> block_rsv, orig_rsv and block_rsv, which means we may lose our
>>> first block_rsv after two joining transaction operations:
>>>
>>> 1) btrfs_start_transaction
>>>       trans->block_rsv = A
>>>
>>> 2) btrfs_join_transaction
>>>       trans->orig_rsv = trans->block_rsv; ---> orig_rsv is now A
>>>       trans->block_rsv = B
>>>
>>> 3) btrfs_join_transaction
>>>       trans->orig_rsv = trans->block_rsv; ---> orig_rsv is now B
>>>       trans->block_rsv = C
>>>    ...
>>>
>>
>> I'd like to see the actual stack trace where this happens, because I don't think
>> it can happen.  And if it is we need to look at that specific case and adjust it
>> as necessary and not add a bunch of kmallocs just to track the block_rsv,
>> because frankly it's not that big of a deal, it was just put into place in case
>> somebody wasn't expecting a call they made to start another transaction and
>> reset the block_rsv, which I don't actually think happens anywhere.  So NAK on
>> this patch, give me more information so I can figure out the right way to deal
>> with this.  Thanks,
>>
> 
> Fine, please run xfstests 068 till it hits a BUG_ON inside either btrfs_delete_delayed_dir_index or
> btrfs_insert_delayed_dir_index.
> 
> What I saw is that the orig_rsv and block_rsv is both delalloc_block_rsv, which is already lack of space.
> 

and trans->use_count has been 3.