Re: [PATCH v2 RFC] f2fs: fix to force keeping write barrier for strict fsync mode

[Jaegeuk Kim]

On 07/14, Chao Yu wrote:
> On 2021/7/14 10:19, Jaegeuk Kim wrote:
> > On 07/14, Chao Yu wrote:
> > > On 2021/7/14 7:34, Jaegeuk Kim wrote:
> > > > On 07/13, Chao Yu wrote:
> > > > > On 2021/7/8 1:48, Jaegeuk Kim wrote:
> > > > > > On 07/02, Chao Yu wrote:
> > > > > > > On 2021/7/2 9:32, Jaegeuk Kim wrote:
> > > > > > > > On 07/02, Chao Yu wrote:
> > > > > > > > > On 2021/7/2 1:10, Jaegeuk Kim wrote:
> > > > > > > > > > On 06/01, Chao Yu wrote:
> > > > > > > > > > > [1] https://www.mail-archive.com/linux-f2fs-devel@xxxxxxxxxxxxxxxxxxxxx/msg15126.html
> > > > > > > > > > > 
> > > > > > > > > > > As [1] reported, if lower device doesn't support write barrier, in below
> > > > > > > > > > > case:
> > > > > > > > > > > 
> > > > > > > > > > > - write page #0; persist
> > > > > > > > > > > - overwrite page #0
> > > > > > > > > > > - fsync
> > > > > > > > > > >       - write data page #0 OPU into device's cache
> > > > > > > > > > >       - write inode page into device's cache
> > > > > > > > > > >       - issue flush
> > > > > > > > > > 
> > > > > > > > > > Well, we have preflush for node writes, so I don't think this is the case.
> > > > > > > > > > 
> > > > > > > > > >       fio.op_flags |= REQ_PREFLUSH | REQ_FUA;
> > > > > > > > > 
> > > > > > > > > This is only used for atomic write case, right?
> > > > > > > > > 
> > > > > > > > > I mean the common case which is called from f2fs_issue_flush() in
> > > > > > > > > f2fs_do_sync_file().
> > > > > > > > 
> > > > > > > > How about adding PREFLUSH when writing node blocks aligned to the above set?
> > > > > > > 
> > > > > > > You mean implementation like v1 as below?
> > > > > > > 
> > > > > > > https://lore.kernel.org/linux-f2fs-devel/20200120100045.70210-1-yuchao0@xxxxxxxxxx/
> > > > > > 
> > > > > > Yea, I think so. :P
> > > > > 
> > > > > I prefer v2, we may have several schemes to improve performance with v2, e.g.
> > > > > - use inplace IO to avoid newly added preflush
> > > > > - use flush_merge option to avoid redundant preflush
> > > > > - if lower device supports barrier IO, we can avoid newly added preflush
> > > > 
> > > > Doesn't v2 give one more flush than v1? Why do you want to take worse one and
> > > 
> > > FUA implies an extra preflush command or similar mechanism in lower device to keep data
> > > in bio being persistent before this command's completion.
> > > 
> > > Also if lower device doesn't support FUA natively, block layer turns it into an empty
> > > PREFLUSH command.
> > > 
> > > So, it's hard to say which one will win the benchmark game, maybe we need some
> > > performance data before making the choice, but you know, it depends on device's
> > > character.
> > 
> > I was looking at # of bios.
> > 
> > > 
> > > > try to improve back? Not clear the benefit on v2.
> > > 
> > > Well, if user suffer and complain performance regression with v1, any plan to improve it?
> > > 
> > > I just thought about plan B/C/D for no matter v1 or v2.
> > 
> > I assumed you wanted v2 since it might be used for B/C/D improvements. But, it
> > seems it wasn't. My point is to save one bio, but piggyback the flag to the
> > device driver.
> 
> I doubt the conclusion...but it needs to get some data to prove it.
> 
> I think the right way is merging v1 now to fix the bug firstly, and let me do
> the comparison on them a little bit later to see whether we need another
> implementation... thoughts?

Chao, could you please post v1 with an updated description?

> 
> Thanks,
> 
> > 
> > > 
> > > Thanks,
> > > 
> > > > 
> > > > > 
> > > > > Thanks,
> > > > > 
> > > > > > 
> > > > > > > 
> > > > > > > Thanks,
> > > > > > > 
> > > > > > > > 
> > > > > > > > > 
> > > > > > > > > And please see do_checkpoint(), we call f2fs_flush_device_cache() and
> > > > > > > > > commit_checkpoint() separately to keep persistence order of CP datas.
> > > > > > > > > 
> > > > > > > > > See commit 46706d5917f4 ("f2fs: flush cp pack except cp pack 2 page at first")
> > > > > > > > > for details.
> > > > > > > > > 
> > > > > > > > > Thanks,
> > > > > > > > > 
> > > > > > > > > > 
> > > > > > > > > > > 
> > > > > > > > > > > If SPO is triggered during flush command, inode page can be persisted
> > > > > > > > > > > before data page #0, so that after recovery, inode page can be recovered
> > > > > > > > > > > with new physical block address of data page #0, however there may
> > > > > > > > > > > contains dummy data in new physical block address.
> > > > > > > > > > > 
> > > > > > > > > > > Then what user will see is: after overwrite & fsync + SPO, old data in
> > > > > > > > > > > file was corrupted, if any user do care about such case, we can suggest
> > > > > > > > > > > user to use STRICT fsync mode, in this mode, we will force to trigger
> > > > > > > > > > > preflush command to persist data in device cache in prior to node
> > > > > > > > > > > writeback, it avoids potential data corruption during fsync().
> > > > > > > > > > > 
> > > > > > > > > > > Signed-off-by: Chao Yu <yuchao0@xxxxxxxxxx>
> > > > > > > > > > > ---
> > > > > > > > > > > v2:
> > > > > > > > > > > - fix this by adding additional preflush command rather than using
> > > > > > > > > > > atomic write flow.
> > > > > > > > > > >       fs/f2fs/file.c | 14 ++++++++++++++
> > > > > > > > > > >       1 file changed, 14 insertions(+)
> > > > > > > > > > > 
> > > > > > > > > > > diff --git a/fs/f2fs/file.c b/fs/f2fs/file.c
> > > > > > > > > > > index 7d5311d54f63..238ca2a733ac 100644
> > > > > > > > > > > --- a/fs/f2fs/file.c
> > > > > > > > > > > +++ b/fs/f2fs/file.c
> > > > > > > > > > > @@ -301,6 +301,20 @@ static int f2fs_do_sync_file(struct file *file, loff_t start, loff_t end,
> > > > > > > > > > >       				f2fs_exist_written_data(sbi, ino, UPDATE_INO))
> > > > > > > > > > >       			goto flush_out;
> > > > > > > > > > >       		goto out;
> > > > > > > > > > > +	} else {
> > > > > > > > > > > +		/*
> > > > > > > > > > > +		 * for OPU case, during fsync(), node can be persisted before
> > > > > > > > > > > +		 * data when lower device doesn't support write barrier, result
> > > > > > > > > > > +		 * in data corruption after SPO.
> > > > > > > > > > > +		 * So for strict fsync mode, force to trigger preflush to keep
> > > > > > > > > > > +		 * data/node write order to avoid potential data corruption.
> > > > > > > > > > > +		 */
> > > > > > > > > > > +		if (F2FS_OPTION(sbi).fsync_mode == FSYNC_MODE_STRICT &&
> > > > > > > > > > > +								!atomic) {
> > > > > > > > > > > +			ret = f2fs_issue_flush(sbi, inode->i_ino);
> > > > > > > > > > > +			if (ret)
> > > > > > > > > > > +				goto out;
> > > > > > > > > > > +		}
> > > > > > > > > > >       	}
> > > > > > > > > > >       go_write:
> > > > > > > > > > >       	/*
> > > > > > > > > > > -- 
> > > > > > > > > > > 2.29.2