Re: [PATCH V4] f2fs: Avoid double lock for cp_rwsem during checkpoint

[Chao Yu]

On 2020/5/9 0:10, Jaegeuk Kim wrote:
> Hi Sayali,
> 
> In order to address the perf regression, how about this?
> 
>>From 48418af635884803ffb35972df7958a2e6649322 Mon Sep 17 00:00:00 2001
> From: Jaegeuk Kim <jaegeuk@xxxxxxxxxx>
> Date: Fri, 8 May 2020 09:08:37 -0700
> Subject: [PATCH] f2fs: avoid double lock for cp_rwsem during checkpoint
> 
> There could be a scenario where f2fs_sync_node_pages gets
> called during checkpoint, which in turn tries to flush
> inline data and calls iput(). This results in deadlock as
> iput() tries to hold cp_rwsem, which is already held at the
> beginning by checkpoint->block_operations().
> 
> Call stack :
> 
> Thread A		Thread B
> f2fs_write_checkpoint()
> - block_operations(sbi)
>  - f2fs_lock_all(sbi);
>   - down_write(&sbi->cp_rwsem);
> 
>                         - open()
>                          - igrab()
>                         - write() write inline data
>                         - unlink()
> - f2fs_sync_node_pages()
>  - if (is_inline_node(page))
>   - flush_inline_data()
>    - ilookup()
>      page = f2fs_pagecache_get_page()
>      if (!page)
>       goto iput_out;
>      iput_out:
> 			-close()
> 			-iput()
>        iput(inode);
>        - f2fs_evict_inode()
>         - f2fs_truncate_blocks()
>          - f2fs_lock_op()
>            - down_read(&sbi->cp_rwsem);
> 
> Fixes: 2049d4fcb057 ("f2fs: avoid multiple node page writes due to inline_data")
> Signed-off-by: Sayali Lokhande <sayalil@xxxxxxxxxxxxxx>
> Signed-off-by: Jaegeuk Kim <jaegeuk@xxxxxxxxxx>
> ---
>  fs/f2fs/node.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
> 
> diff --git a/fs/f2fs/node.c b/fs/f2fs/node.c
> index 1db8cabf727ef..626d7daca09de 100644
> --- a/fs/f2fs/node.c
> +++ b/fs/f2fs/node.c
> @@ -1870,8 +1870,8 @@ int f2fs_sync_node_pages(struct f2fs_sb_info *sbi,
>  				goto continue_unlock;
>  			}
>  
> -			/* flush inline_data */
> -			if (is_inline_node(page)) {
> +			/* flush inline_data, if it's not sync path. */
> +			if (do_balance && is_inline_node(page)) {

IIRC, this flow was designed to avoid running out of free space issue
during checkpoint:

2049d4fcb057 ("f2fs: avoid multiple node page writes due to inline_data")

The sceanrio is:
1. create fully node blocks
2. flush node blocks
3. write inline_data for all the node blocks again
4. flush node blocks redundantly

I guess this may cause failing one case of fstest.


Since block_operations->f2fs_sync_inode_meta has synced inode cache to
inode page, so in block_operations->f2fs_sync_node_pages, could we
check nlink before flush_inline_data():

if (is_inline_node(page)) {
	if (IS_INODE(page) && raw_inode_page->i_links) {
		flush_inline_data()
	}
}


>  				clear_inline_node(page);
>  				unlock_page(page);
>  				flush_inline_data(sbi, ino_of_node(page));
>