Re: [PATCH -next v2] ext4: fix bug_on in ext4_writepages

[Jan Kara <jack@suse.cz>]

On Tue 10-05-22 18:02:28, 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))
> 
> The root cause of this issue is we destory inline data until call ext4_writepages
> under delay allocation mode. But there maybe already covert from inline to extent.
> To solved this issue, we call filemap_flush firstly.
> 
> Signed-off-by: Ye Bin <yebin10@huawei.com>
> ---
>  fs/ext4/inline.c | 8 ++++++++
>  1 file changed, 8 insertions(+)
> 
> diff --git a/fs/ext4/inline.c b/fs/ext4/inline.c
> index 6d253edebf9f..130ed5d83734 100644
> --- a/fs/ext4/inline.c
> +++ b/fs/ext4/inline.c
> @@ -2002,6 +2002,14 @@ int ext4_convert_inline_data(struct inode *inode)
>  	if (!ext4_has_inline_data(inode)) {
>  		ext4_clear_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA);
>  		return 0;
> +	} else if (test_opt(inode->i_sb, DELALLOC) && !S_ISDIR(inode->i_mode)) {
> +		error = filemap_flush(inode->i_mapping);

This is actually an interesting option and I kind of like it but shouldn't
we restrict this to the situation when EXT4_STATE_MAY_INLINE_DATA is clear?
Otherwise we would be writing out inline data to the inode unnecessarily
for each ext4_convert_inline_data() call.

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