Re: [f2fs-dev] [PATCH v2] f2fs: avoid sb_start_intwrite during eviction

[Chao Yu <chao@kernel.org>]

On 2022/3/1 12:48, Jaegeuk Kim wrote:
> 1. waiting for f2fs_evict_inode
> [ 5560.043945]  __wait_on_freeing_inode+0xac/0xf0
> [ 5560.045540]  ? var_wake_function+0x30/0x30
> [ 5560.047036]  find_inode_fast+0x6d/0xc0
> [ 5560.048473]  iget_locked+0x79/0x230
> [ 5560.049933]  f2fs_iget+0x27/0x1200 [f2fs]
> [ 5560.051496]  f2fs_lookup+0x18c/0x3e0 [f2fs]
> [ 5560.053069]  __lookup_slow+0x84/0x150
> [ 5560.054503]  walk_component+0x141/0x1b0
> [ 5560.055938]  link_path_walk.part.0+0x23b/0x360
> [ 5560.057541]  ? end_bio_bh_io_sync+0x37/0x50
> [ 5560.059086]  path_parentat+0x3c/0x90
> [ 5560.060492]  filename_parentat+0xd7/0x1e0
> [ 5560.062002]  ? blk_mq_free_request+0x127/0x150
> [ 5560.063576]  do_renameat2+0xc1/0x5b0
>   --> sb_start_write(m->mnt_sb); ->  __sb_start_write(sb, SB_FREEZE_WRITE);
> 
> [ 5560.064999]  ? __check_object_size+0x13f/0x150
> [ 5560.066559]  ? strncpy_from_user+0x44/0x150
> [ 5560.068038]  ? getname_flags.part.0+0x4c/0x1b0
> [ 5560.069617]  __x64_sys_renameat2+0x51/0x60
> 
> 2. waiting for sb_start_intwrite -> __sb_start_write(sb, SB_FREEZE_FS);

It's still not clear that why __sb_start_write(sb, SB_FREEZE_FS) will be blocked,
as SB_FREEZE_FS and SB_FREEZE_WRITE points to different locks.

Thread A				Thread B				Thread C
- rename
  - sb_start_write
   - __sb_start_write(SB_FREEZE_WRITE)
...
      - f2fs_lookup
...
        - __wait_on_freeing_inode
					- drop_slab
					 - prune_icache_sb
					  - inode_lru_isolate
					   :inode->i_state |= I_FREEING
										- Is there any flow that it has already held
										 SB_FREEZE_FS and try to lock SB_FREEZE_WRITE?
					   - f2fs_evict_inode
					    - __sb_start_write(SB_FREEZE_FS)

Thanks,

> 
> [ 5560.152447]  percpu_rwsem_wait+0xaf/0x160
> [ 5560.154000]  ? percpu_down_write+0xd0/0xd0
> [ 5560.155498]  __percpu_down_read+0x4e/0x60
> [ 5560.157000]  f2fs_evict_inode+0x5a3/0x610 [f2fs]
> [ 5560.158648]  ? var_wake_function+0x30/0x30
> [ 5560.160341]  evict+0xd2/0x180
> [ 5560.161728]  prune_icache_sb+0x81/0xb0
>   --> inode_lru_isolate() -> inode->i_state |= I_FREEING;
> 
> [ 5560.163179]  super_cache_scan+0x169/0x1f0
> [ 5560.164675]  do_shrink_slab+0x145/0x2b0
> [ 5560.166121]  shrink_slab+0x186/0x2d0
> [ 5560.167481]  drop_slab_node+0x4a/0x90
> [ 5560.168876]  drop_slab+0x3e/0x80
> [ 5560.170178]  drop_caches_sysctl_handler+0x75/0x90
> [ 5560.171761]  proc_sys_call_handler+0x149/0x280
> [ 5560.173328]  proc_sys_write+0x13/0x20
> [ 5560.174667]  new_sync_write+0x117/0x1b0
> [ 5560.176120]  vfs_write+0x1d5/0x270
> [ 5560.177409]  ksys_write+0x67/0xe0
> 
> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
> ---
>   Note, I found this call stack.
> 
>   fs/f2fs/inode.c | 2 --
>   1 file changed, 2 deletions(-)
> 
> diff --git a/fs/f2fs/inode.c b/fs/f2fs/inode.c
> index ab8e0c06c78c..882db4bd917b 100644
> --- a/fs/f2fs/inode.c
> +++ b/fs/f2fs/inode.c
> @@ -778,7 +778,6 @@ void f2fs_evict_inode(struct inode *inode)
>   	f2fs_remove_ino_entry(sbi, inode->i_ino, UPDATE_INO);
>   	f2fs_remove_ino_entry(sbi, inode->i_ino, FLUSH_INO);
>   
> -	sb_start_intwrite(inode->i_sb);
>   	set_inode_flag(inode, FI_NO_ALLOC);
>   	i_size_write(inode, 0);
>   retry:
> @@ -809,7 +808,6 @@ void f2fs_evict_inode(struct inode *inode)
>   		if (dquot_initialize_needed(inode))
>   			set_sbi_flag(sbi, SBI_QUOTA_NEED_REPAIR);
>   	}
> -	sb_end_intwrite(inode->i_sb);
>   no_delete:
>   	dquot_drop(inode);
>