Re: Data lost in Android app for not write new checkpoint

[He YunLei <heyunlei@huawei.com>]

On 2015/7/31 18:49, Chao Yu wrote:
> Hi Bintian,
>
>> -----Original Message-----
>> From: He YunLei [mailto:heyunlei@huawei.com]
>> Sent: Friday, July 31, 2015 10:29 AM
>> To: linux-f2fs-devel@lists.sourceforge.net; Jaegeuk Kim
>> Cc: Chao Yu; cm224.lee@samsung.com; Bintian
>> Subject: [f2fs-dev] Data lost in Android app for not write new checkpoint
>>
>> Hi all,
>> 	Recently I did some test with f2fs on my Android phone, and found a problem
>> which I didn't know how to tackle it.
>> 	I use my Android phone with /data partition formatted  by mkfs.f2fs. When the
>> phone just started, I check the f2fs status by reading the file /sys/kernel/debug/f2fs/status
>> in debugfs.
>>
>> CP calls: 10
>> GC calls: 19 (BG: 19)
>>     - data segments : 19 (19)
>>     - node segments : 0 (0)
>>
>> 	We can see /data partition has done 10 times write_checkpoint since f2fs is mounted
>> on the phone, it also has triggered 19 times background GC.
>>
>> ******
>>
>> Here I took some photos consecutively, and check the file /sys/kernel/debug/f2fs/status again
>>
>> ******
>>
>> CP calls: 10
>> GC calls: 20 (BG: 20)
>>     - data segments : 20 (20)
>>     - node segments : 0 (0)
>>
>> 	there is no change in CP calls number and background GC doesn't write new checkpoint.
>> if then a sudden power failure or system crash occur, the photos will be lost when the phone
>> restart, and a sync before crash will avoid the data lost.
>> 	I think this problem is bad for user experience of using Android phone with f2fs.
>> How do we deal with such situation? I wish you and other developers in this list could help
>> me in a correct way.
>
> IMO, it's better to figure out whether this is a bug of f2fs first or not.
>
> You can enable some traces in f2fs to see whether fsync is called or not.
>
> enable trace by:
> echo 1 > /sys/kernel/debug/tracing/events/f2fs/f2fs_sync_file_enter/enable
> echo 1 > /sys/kernel/debug/tracing/events/f2fs/f2fs_sync_file_exit/enable
> print trace by:
> cat /sys/kernel/debug/tracing/trace
>
> If fsync is not be called, I think in ext4 there must be the same problem,
> but I guess fortunately journal commit thread save its data since it commit
> transaction per 5 second by default. You can try to configure (commit=nrsec)
> it with larger value for verification the issue with ext4 filesystem.
>

I enable the event xxx_sync_file_enter both in f2fs and ext4, and find neither of
them was triggered by photo files.

Then I try f2fs_writepages and ext4_da_write_pages:

    ino     file_name
				
    65573   IMG_20150804_031619.jpg
    65575   IMG_20150804_031619_1.jpg
    65576   IMG_20150804_031620.jpg
    65577   IMG_20150804_031620_1.jpg

  ext4_da_write_pages: dev 259,0 ino 65573 b_blocknr 0 b_size 0 b_state 0x0000 first_page 0 io_done 0 pages_written 0 sync_mode 0
  ext4_da_write_pages: dev 259,0 ino 65575 b_blocknr 0 b_size 2408448 b_state 0x0221 first_page 0 io_done 1 pages_written 588 sync_mode 0
  ext4_da_write_pages: dev 259,0 ino 65575 b_blocknr 0 b_size 0 b_state 0x0000 first_page 0 io_done 0 pages_written 0 sync_mode 0
  ext4_da_write_pages: dev 259,0 ino 65576 b_blocknr 0 b_size 2428928 b_state 0x0221 first_page 0 io_done 1 pages_written 593 sync_mode 0
  ext4_da_write_pages: dev 259,0 ino 65576 b_blocknr 0 b_size 0 b_state 0x0000 first_page 0 io_done 0 pages_written 0 sync_mode 0
  ext4_da_write_pages: dev 259,0 ino 65577 b_blocknr 0 b_size 2383872 b_state 0x0221 first_page 0 io_done 1 pages_written 582 sync_mode 0
  ext4_da_write_pages: dev 259,0 ino 65577 b_blocknr 0 b_size 0 b_state 0x0000 first_page 0 io_done 0 pages_written 0 sync_mode 0

f2fs_writepages doesn't appear in the test of f2fs

I also try modify commit=300(default 5), but it doesn't work. Maybe somewhere else in ext4
launch the ext4_da_write_pages operation.

At the end, I try to mount f2fs with disable_roll_forward, when system reboot, the f2fs is inconsistent,
there are several failed check items in fsck.

Thanks,
He

> As a quick thought, maybe we can add one commit data thread, periodically
> writebacking user data written by user previously, then do checkpoint for
> persistence.
>
> So by this way, at most, we just lose our data for last configured time of
> commit period.
>
> Thanks,
>
>>
>> Thanks,
>> He
>
>
>
> .
>


------------------------------------------------------------------------------