Re: [f2fs-dev] [PATCH] f2fs: check if inode's state is dirty or not before skip fsync

[Jaegeuk Kim <jaegeuk@kernel.org>]

On Fri, Dec 05, 2014 at 11:10:15AM +0900, Changman Lee wrote:
> On Thu, Dec 04, 2014 at 04:58:29PM -0800, Jaegeuk Kim wrote:
> > On Wed, Dec 03, 2014 at 10:46:38AM +0900, Changman Lee wrote:
> > > Hi Jaegeuk,
> > > 
> > > Thanks for explanation.
> > > 
> > > On Tue, Dec 02, 2014 at 11:42:19AM -0800, Jaegeuk Kim wrote:
> > > > On Tue, Dec 02, 2014 at 01:21:31PM +0900, Changman Lee wrote:
> > > > > Hi,
> > > > > 
> > > > > f2fs_dirty_inode just set fi->flag as FI_DIRTY_INODE not to
> > > > > call update_inode_page. Instead, we do it when f2fs_write_indoe is called.
> > > > > Do you have any reason to do like this?
> > > > 
> > > > Actually, I'd like to use inode caches instead of dirty node pages as much as
> > > > possible to mitigate memory pressure as well as reduce node page writes.
> > > > But, the reality is that f2fs triggers update_inode_page frequently, since some
> > > > inode information like i_blocks and i_links should be recovered consistently
> > > > from sudden power-cuts.
> > > 
> > > I got it. No objection.
> > > 
> > > > 
> > > > > How about move update_inode_page from write_inode to dirty_inode?
> > > > > And we can update inode page when mark_inode_dirty or
> > > > > mark_inode_dirty_sync is called. Then, we control write I/O in
> > > > > write_inode according to wbc->sync_mode.
> > > > 
> > > > What do you mean controlling write I/O in write_inode?
> > > > The write_inode does not trigger any I/Os.
> > > > We're controlling node page writes by f2fs_write_node_pages.
> > > 
> > > Sorry, it's not enough for my explanation.
> > > At __writeback_single_inode, it calls write_inode if inode is dirty.
> > > And at ext4_write_inode and btrfs_write_inode, they issue write
> > > according to wbc->sync_mode. However, current f2fs doesn't issue any
> > > write i/o. Could you review it?
> > 
> > Hi,
> > 
> > Well, I'm not quite sure that f2fs should do this.
> > In terms of recovery, we don't need to do this.
> > 
> > > 
> > > > 
> > > > Anyway, if we call update_inode_page in mark_inode_dirty, f2fs would suffer from
> > > > a lot of dirty node pages.
> > > 
> > > Got it. But I think we should write dirty node after
> > > update_inode_page in write_inode if wbc->sync_mode == WB_SYNC_ALL.
> > 
> > Why do we have to do this?
> > Again, there is no problem wrt recovery, but that causes unnecessary IOs.
> > 
> > > 
> > > 
> > > Finally, I have one more question.
> > > At f2fs_sync_file, in the case of need_cp is true and file_wrong_pino
> > > f2fs calls write_inode. But the inode isn't written back. Is it okay?
> > > Could you elaborate on it?
> > 
> > No problem. That pino will be used only for fsynced inodes after checkpoint.
> 
> I got it. My concern was started from this. If there is no problem,
> I think current f2fs_write_inode is also no problem.
> Thanks Jaegeuk.
> 
> Then, let's merge your suggestion below.

Please, write and resubmit patch v2.
One additional comment is that conditions in the patch should have considered
datasync like this.

	/* if the inode is dirty, let's recover all the time */
	if (!datasync && is_inode_flag_set(fi, FI_DIRTY_INODE)) {
		update_inode_page(inode);
		goto go_write;
	}

}
>
> 
> Lastly, I have curiosity related to write node; APPEND or UPDATE.
> Before fsync is called, isn't there any possiblity to be changed to APPEND from
> UPDATE. If so, we might lost recovery info.

No, APPEND and UPDATE are orthogonal.

Thanks,

> I think we'd better check if there is a situation.
> 
> Regards,
> Changman
> 
> > 
> > Thanks,
> > 
> > > 
> > > Thanks,
> > > 
> > > > 
> > > > Thanks,
> > > > 
> > > > > Could you consider this once?
> > > > > 
> > > > > Thanks,
> > > > > 
> > > > > On Mon, Dec 01, 2014 at 02:52:57PM -0800, Jaegeuk Kim wrote:
> > > > > > On Mon, Dec 01, 2014 at 04:05:20PM +0900, Changman Lee wrote:
> > > > > > > It makes sense to check inode's state than check if
> > > > > > > inode page is dirty or not.
> > > > > > 
> > > > > > Nice catch.
> > > > > > However, we should leave the original condition, since write_inode can be called
> > > > > > in prior to this fsync call.
> > > > > > And, this is not a proper fix, since it still can skip to write its inode page. 
> > > > > > 
> > > > > > How about this one?
> > > > > > 
> > > > > > diff --git a/fs/f2fs/file.c b/fs/f2fs/file.c
> > > > > > index 146e58a..6690599 100644
> > > > > > --- a/fs/f2fs/file.c
> > > > > > +++ b/fs/f2fs/file.c
> > > > > > @@ -168,6 +168,12 @@ int f2fs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
> > > > > >  		return ret;
> > > > > >  	}
> > > > > >  
> > > > > > +	/* if the inode is dirty, let's recover all the time */
> > > > > > +	if (is_inode_flag_set(fi, FI_DIRTY_INODE)) {
> > > > > > +		update_inode_page(inode);
> > > > > > +		goto go_write;
> > > > > > +	}
> > > > > > +
> > > > > >  	/*
> > > > > >  	 * if there is no written data, don't waste time to write recovery info.
> > > > > >  	 */
> > > > > > -- 
> > > > > > 2.1.1
> > > > > > 
> > > > > > > 
> > > > > > > Signed-off-by: Changman Lee <cm224.lee@samsung.com>
> > > > > > > ---
> > > > > > >  fs/f2fs/file.c | 7 ++-----
> > > > > > >  1 file changed, 2 insertions(+), 5 deletions(-)
> > > > > > > 
> > > > > > > diff --git a/fs/f2fs/file.c b/fs/f2fs/file.c
> > > > > > > index 7c2ec3e..0c5ae87 100644
> > > > > > > --- a/fs/f2fs/file.c
> > > > > > > +++ b/fs/f2fs/file.c
> > > > > > > @@ -173,14 +173,11 @@ int f2fs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
> > > > > > >  	 */
> > > > > > >  	if (!is_inode_flag_set(fi, FI_APPEND_WRITE) &&
> > > > > > >  			!exist_written_data(sbi, ino, APPEND_INO)) {
> > > > > > > -		struct page *i = find_get_page(NODE_MAPPING(sbi), ino);
> > > > > > >  
> > > > > > >  		/* But we need to avoid that there are some inode updates */
> > > > > > > -		if ((i && PageDirty(i)) || need_inode_block_update(sbi, ino)) {
> > > > > > > -			f2fs_put_page(i, 0);
> > > > > > > +		if (is_inode_flag_set(fi, FI_DIRTY_INODE) ||
> > > > > > > +					need_inode_block_update(sbi, ino))
> > > > > > >  			goto go_write;
> > > > > > > -		}
> > > > > > > -		f2fs_put_page(i, 0);
> > > > > > >  
> > > > > > >  		if (is_inode_flag_set(fi, FI_UPDATE_WRITE) ||
> > > > > > >  				exist_written_data(sbi, ino, UPDATE_INO))
> > > > > > > -- 
> > > > > > > 1.9.1
> > > > > > > 
> > > > > > > 
> > > > > > > ------------------------------------------------------------------------------
> > > > > > > Download BIRT iHub F-Type - The Free Enterprise-Grade BIRT Server
> > > > > > > from Actuate! Instantly Supercharge Your Business Reports and Dashboards
> > > > > > > with Interactivity, Sharing, Native Excel Exports, App Integration & more
> > > > > > > Get technology previously reserved for billion-dollar corporations, FREE
> > > > > > > http://pubads.g.doubleclick.net/gampad/clk?id=157005751&iu=/4140/ostg.clktrk
> > > > > > > _______________________________________________
> > > > > > > Linux-f2fs-devel mailing list
> > > > > > > Linux-f2fs-devel@lists.sourceforge.net
> > > > > > > https://lists.sourceforge.net/lists/listinfo/linux-f2fs-devel