On 08/20/2014 06:52 AM, Miao Xie wrote:
> On Tue, 19 Aug 2014 10:58:09 -0400, Chris Mason wrote:
>> On 08/19/2014 10:23 AM, David Sterba wrote:
>>> On Tue, Aug 19, 2014 at 07:58:20PM +0800, Fengguang Wu wrote:
>>>> We noticed an xfstests failure on commit
>>>>
>>>> 8d875f95da43c6a8f18f77869f2ef26e9594fecc ("btrfs: disable strict file flushes for renames and truncates")
>>>>
>>>> It's 100% reproducible in the 5 test runs.
>>>
>>> Same here, different mkfs configurations.
>>>
>>> generic/226 28s ...    [16:11:52] [16:12:55] - output mismatch (see /root/xfstests/results//generic/226.out.bad)
>>>     --- tests/generic/226.out   2013-05-29 17:16:03.000000000 +0200
>>>     +++ /root/xfstests/results//generic/226.out.bad     2014-08-19 16:12:55.000000000 +0200
>>>     @@ -1,6 +1,8 @@
>>>      QA output created by 226
>>>      --> mkfs 256m filesystem
>>>      --> 16 buffered 64m writes in a loop
>>>     -1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
>>>     +1 2 3 4 pwrite64: No space left on device
>>>     +5 6 7 8 9 10 11 12 pwrite64: No space left on device
>>>     +13 14 15 16
>>>
>>> enospc on a small filesystem (256M)
>>
>> I'm calling filemap flush more often, but otherwise everything else is
>> the same.  I'll take a look.
> 
> I found the problem is caused by the following function:
> 
> int btrfs_release_file(struct inode *inode, struct file *filp)
> {
> 	...
> 	filemap_flush(inode->i_mapping);
> 	return 0;
> }
> 
> I don't think we need flush file at most situation. Ext4 flushes the file only
> after someone truncate the file to be zero-length, I don't know the real reason
> why ext4 flush the file only after the file is truncated, someone said it is to
> reduce the risk that the users find a zero-length file after a crash, which happens
> after truncate-write-close process.
> 
> If we change btrfs_release_file by ext4's implementation, both the failure of
> xfstests's generic/226  and performance regression can be fixed.
> 

You're completely right, my original had more checks here and I stripped
them out by accident.  Fixing, thanks!

-chris