[Ocfs2-devel] [PATCH 1/2] ocfs2: fix jbd2 assertion in defragment path

[Joseph Qi]

Hi,

Sorry for the late response since I was busy on other things...

On 5/21/22 6:14 PM, Heming Zhao wrote:
> defragfs triggered jbd2 report ASSERT when working.
> 
> code path:
> 
> ocfs2_ioctl_move_extents
>  ocfs2_move_extents
>   __ocfs2_move_extents_range
>    ocfs2_defrag_extent
>     __ocfs2_move_extent
>      + ocfs2_journal_access_di
>      + ocfs2_split_extent  //[1]
>      + ocfs2_journal_dirty //crash
> 
> [1]: ocfs2_split_extent called ocfs2_extend_trans, which committed
> dirty buffers then restarted the transaction. The changed transaction
> triggered jbd2 ASSERT.
> 
> crash stacks:
> 
> PID: 11297  TASK: ffff974a676dcd00  CPU: 67  COMMAND: "defragfs.ocfs2"
>  #0 [ffffb25d8dad3900] machine_kexec at ffffffff8386fe01
>  #1 [ffffb25d8dad3958] __crash_kexec at ffffffff8395959d
>  #2 [ffffb25d8dad3a20] crash_kexec at ffffffff8395a45d
>  #3 [ffffb25d8dad3a38] oops_end at ffffffff83836d3f
>  #4 [ffffb25d8dad3a58] do_trap at ffffffff83833205
>  #5 [ffffb25d8dad3aa0] do_invalid_op at ffffffff83833aa6
>  #6 [ffffb25d8dad3ac0] invalid_op at ffffffff84200d18
>     [exception RIP: jbd2_journal_dirty_metadata+0x2ba]
>     RIP: ffffffffc09ca54a  RSP: ffffb25d8dad3b70  RFLAGS: 00010207
>     RAX: 0000000000000000  RBX: ffff9706eedc5248  RCX: 0000000000000000
>     RDX: 0000000000000001  RSI: ffff97337029ea28  RDI: ffff9706eedc5250
>     RBP: ffff9703c3520200   R8: 000000000f46b0b2   R9: 0000000000000000
>     R10: 0000000000000001  R11: 00000001000000fe  R12: ffff97337029ea28
>     R13: 0000000000000000  R14: ffff9703de59bf60  R15: ffff9706eedc5250
>     ORIG_RAX: ffffffffffffffff  CS: 0010  SS: 0018
>  #7 [ffffb25d8dad3ba8] ocfs2_journal_dirty at ffffffffc137fb95 [ocfs2]
>  #8 [ffffb25d8dad3be8] __ocfs2_move_extent at ffffffffc139a950 [ocfs2]
>  #9 [ffffb25d8dad3c80] ocfs2_defrag_extent at ffffffffc139b2d2 [ocfs2]
> 
> The fix method is simple, ocfs2_split_extent includes three paths. all
> the paths already have journal access/dirty pair. We only need to
> remove journal access/dirty from __ocfs2_move_extent.
> 

I am not sure what you mean by "all three paths have journal access/
dirty pair".

It seems we can't do it in your way, as journal access has different
ocfs2_trigger with different offset, e.g. dinode is different from
extent block.

Or we may reserve enough credits to resolve this issue? 

Thanks,
Joseph