Re: [PATCH -next] ext4: fix bug_on in ext4_writepages

[Jan Kara <jack@suse.cz>]

On Fri 06-05-22 20:48:44, yebin wrote:
> On 2022/5/6 16:50, Jan Kara wrote:
> > On Fri 06-05-22 09:27:25, yebin wrote:
> > > On 2022/5/5 23:47, Jan Kara wrote:
> > > > On Thu 05-05-22 21:57:08, Ye Bin wrote:
> > > > > we got issue as follows:
> > > > > EXT4-fs error (device loop0): ext4_mb_generate_buddy:1141: group 0, block bitmap and bg descriptor inconsistent: 25 vs 31513 free cls
> > > > > ------------[ cut here ]------------
> > > > > kernel BUG at fs/ext4/inode.c:2708!
> > > > > invalid opcode: 0000 [#1] PREEMPT SMP KASAN PTI
> > > > > CPU: 2 PID: 2147 Comm: rep Not tainted 5.18.0-rc2-next-20220413+ #155
> > > > > RIP: 0010:ext4_writepages+0x1977/0x1c10
> > > > > RSP: 0018:ffff88811d3e7880 EFLAGS: 00010246
> > > > > RAX: 0000000000000000 RBX: 0000000000000001 RCX: ffff88811c098000
> > > > > RDX: 0000000000000000 RSI: ffff88811c098000 RDI: 0000000000000002
> > > > > RBP: ffff888128140f50 R08: ffffffffb1ff6387 R09: 0000000000000000
> > > > > R10: 0000000000000007 R11: ffffed10250281ea R12: 0000000000000001
> > > > > R13: 00000000000000a4 R14: ffff88811d3e7bb8 R15: ffff888128141028
> > > > > FS:  00007f443aed9740(0000) GS:ffff8883aef00000(0000) knlGS:0000000000000000
> > > > > CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> > > > > CR2: 0000000020007200 CR3: 000000011c2a4000 CR4: 00000000000006e0
> > > > > DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
> > > > > DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
> > > > > Call Trace:
> > > > >    <TASK>
> > > > >    do_writepages+0x130/0x3a0
> > > > >    filemap_fdatawrite_wbc+0x83/0xa0
> > > > >    filemap_flush+0xab/0xe0
> > > > >    ext4_alloc_da_blocks+0x51/0x120
> > > > >    __ext4_ioctl+0x1534/0x3210
> > > > >    __x64_sys_ioctl+0x12c/0x170
> > > > >    do_syscall_64+0x3b/0x90
> > > > > 
> > > > > It may happen as follows:
> > > > > 1. write inline_data inode
> > > > > vfs_write
> > > > >     new_sync_write
> > > > >       ext4_file_write_iter
> > > > >         ext4_buffered_write_iter
> > > > >           generic_perform_write
> > > > > 	  ext4_da_write_begin
> > > > > 	    ext4_da_write_inline_data_begin -> If inline data size too
> > > > > 	    small will allocate block to write, then mapping will has
> > > > > 	    dirty page
> > > > > 	    	ext4_da_convert_inline_data_to_extent ->clear EXT4_STATE_MAY_INLINE_DATA
> > > > > 2. fallocate
> > > > > do_vfs_ioctl
> > > > >     ioctl_preallocate
> > > > >       vfs_fallocate
> > > > >         ext4_fallocate
> > > > >           ext4_convert_inline_data
> > > > > 	  ext4_convert_inline_data_nolock
> > > > > 	    ext4_map_blocks -> fail will goto restore data
> > > > > 	    ext4_restore_inline_data
> > > > > 	      ext4_create_inline_data
> > > > > 	      ext4_write_inline_data
> > > > > 	      ext4_set_inode_state -> set inode EXT4_STATE_MAY_INLINE_DATA
> > > > > 3. writepages
> > > > > __ext4_ioctl
> > > > >     ext4_alloc_da_blocks
> > > > >       filemap_flush
> > > > >         filemap_fdatawrite_wbc
> > > > >           do_writepages
> > > > > 	  ext4_writepages
> > > > > 	    if (ext4_has_inline_data(inode))
> > > > > 	      BUG_ON(ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA))
> > > > > 
> > > > > To solved this issue, record origin 'EXT4_STATE_MAY_INLINE_DATA' flag, then pass
> > > > > value to 'ext4_restore_inline_data', 'ext4_restore_inline_data' will
> > > > > decide to if recovery 'EXT4_STATE_MAY_INLINE_DATA' flag according to parameter.
> > > > > 
> > > > > Signed-off-by: Ye Bin <yebin10@huawei.com>
> > > > I think this will get also fixed by a patch from your colleague I've
> > > > reviewed here [1], won't it?
> > > > 
> > > > [1] https://lore.kernel.org/all/20220428165725.mvjh6mx7gr5vekqe@quack3.lan
> > > > 
> > > The issue I fixed is not the same as the isuue my colleague fixed.
> > OK, maybe I've jumped to conclusion too early but the fix I've referenced
> > above will protect ext4_convert_inline_data() in ext4_fallocate() with
> > inode->i_rwsem so I think the race you describe with ext4_da_write_begin()
> > cannot happen. The inline conversion path with be entered either from
> > ext4_da_write_begin() or from ext4_fallocate() but not from both. If I'm
> > missing something, please explain how you think the problem happens with
> > the above fix applied... Thanks!
> > 
> > 								Honza
> It's happen as follows:
> step1:
> ext4_file_write_iter
>   ext4_buffered_write_iter
>     generic_perform_write
>       ext4_da_write_begin
>         if (ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA))
>           ext4_da_write_inline_data_begin
>             ext4_da_convert_inline_data_to_extent
>               __block_write_begin
>               ext4_clear_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA);
>             ->clear EXT4_STATE_MAY_INLINE_DATA flag,  If inline data size
> too
>         small will allocate block to write, then mapping will has dirty page
> 
>       ext4_da_write_end
>         generic_write_end
> 
> step2：
> vfs_fallocate
>   ext4_fallocate
>     ext4_convert_inline_data
>       if (ext4_has_inline_data(inode))  -> This condition is satisfied
>         ext4_convert_inline_data_nolock
>           ext4_map_blocks -> fail will goto restore data
>           ext4_restore_inline_data
>            ext4_create_inline_data
>            ext4_write_inline_data
>            ext4_set_inode_state -> set inode EXT4_STATE_MAY_INLINE_DATA
> step3:
> do_writepages
>   ext4_writepages
>     if (ext4_has_inline_data(inode))
>       BUG_ON(ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA))
> 
> As we call "ext4_destroy_inline_data" to destory inline data when delay
> allocation.  So，if we call fallocate before writepages while
> ext4_map_blocks return failed, will lead to above issue.  This issue does
> not require concurrency conditions.

Aha, I see now. Thanks for explanation and sorry for being a bit dense.
I'll have a look at your original patch.

								Honza
-- 
Jan Kara <jack@suse.com>
SUSE Labs, CR