Re: [PATCH v4] f2fs: add bio cache for IPU

[Jaegeuk Kim]

On 09/02, Chao Yu wrote:
> On 2019/9/1 15:17, Jaegeuk Kim wrote:
> > On 08/31, Chao Yu wrote:
> >> On 2019/2/19 16:15, Chao Yu wrote:
> >>> @@ -1976,10 +2035,13 @@ static int __write_data_page(struct page *page, bool *submitted,
> >>>  	}
> >>>  
> >>>  	unlock_page(page);
> >>> -	if (!S_ISDIR(inode->i_mode) && !IS_NOQUOTA(inode))
> >>> +	if (!S_ISDIR(inode->i_mode) && !IS_NOQUOTA(inode)) {
> >>> +		f2fs_submit_ipu_bio(sbi, bio, page);
> >>>  		f2fs_balance_fs(sbi, need_balance_fs);
> >>> +	}
> >>
> >> Above bio submission was added to avoid below deadlock:
> >>
> >> - __write_data_page
> >>  - f2fs_do_write_data_page
> >>   - set_page_writeback        ---- set writeback flag
> >>   - f2fs_inplace_write_data
> >>  - f2fs_balance_fs
> >>   - f2fs_gc
> >>    - do_garbage_collect
> >>     - gc_data_segment
> >>      - move_data_page
> >>       - f2fs_wait_on_page_writeback
> >>        - wait_on_page_writeback  --- wait writeback
> >>
> >> However, it breaks the merge of IPU IOs, to solve this issue, it looks we need
> >> to add global bio cache for such IPU merge case, then later
> >> f2fs_wait_on_page_writeback can check whether writebacked page is cached or not,
> >> and do the submission if necessary.
> >>
> >> Jaegeuk, any thoughts?
> > 
> > How about calling f2fs_submit_ipu_bio() when we need to do GC in the same
> > context?
> 
> However it also could happen in race case:
> 
> Thread A				Thread B
> - __write_data_page (inode x, page y)
>  - f2fs_do_write_data_page
>   - set_page_writeback        ---- set writeback flag in page y
>   - f2fs_inplace_write_data
>  - f2fs_balance_fs
> 					 - lock gc_mutex
>  - lock gc_mutex
> 					  - f2fs_gc
> 					   - do_garbage_collect
> 					    - gc_data_segment
> 					     - move_data_page
> 					      - f2fs_wait_on_page_writeback
> 					       - wait_on_page_writeback  --- wait writeback of page y
> 
> So it needs a global bio cache for merged IPU pages, how about adding a list to
> link all ipu bios in struct f2fs_bio_info?

Hmm, I can't think of better solution than adding a list. In this case, blk_plug
doesn't work well?

> 
> struct f2fs_bio_info {
> 	....
> 	struct list_head ipu_bio_list;	/* track all ipu bio */
> 	spinlock_t ipu_bio_lock;	/* protect ipu bio list */
> }
> 
> > 
> >>
> >> Thanks,
> > .
> >