Re: [f2fs-dev] [PATCH] f2fs: get parent inode when recovering pino

[Jaegeuk Kim]

On 05/05, Eric Biggers wrote:
> On Tue, May 05, 2020 at 09:58:47AM -0700, Eric Biggers wrote:
> > On Tue, May 05, 2020 at 08:31:39AM -0700, Jaegeuk Kim wrote:
> > > We had to grab the inode before retrieving i_ino.
> > > 
> > > Signed-off-by: Jaegeuk Kim <jaegeuk@xxxxxxxxxx>
> > > ---
> > >  fs/f2fs/file.c | 8 +++++++-
> > >  1 file changed, 7 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/fs/f2fs/file.c b/fs/f2fs/file.c
> > > index a0a4413d6083b..9d4c3e3503567 100644
> > > --- a/fs/f2fs/file.c
> > > +++ b/fs/f2fs/file.c
> > > @@ -168,6 +168,7 @@ static const struct vm_operations_struct f2fs_file_vm_ops = {
> > >  static int get_parent_ino(struct inode *inode, nid_t *pino)
> > >  {
> > >  	struct dentry *dentry;
> > > +	struct inode *parent;
> > >  
> > >  	inode = igrab(inode);
> > >  	dentry = d_find_any_alias(inode);
> > > @@ -175,8 +176,13 @@ static int get_parent_ino(struct inode *inode, nid_t *pino)
> > >  	if (!dentry)
> > >  		return 0;
> > >  
> > > -	*pino = parent_ino(dentry);
> > > +	parent = igrab(d_inode(dentry->d_parent));
> > >  	dput(dentry);
> > > +	if (!parent)
> > > +		return 0;
> > > +
> > > +	*pino = parent->i_ino;
> > > +	iput(parent);
> > >  	return 1;
> > 
> > This doesn't appear to be necessary.  parent_ino() is:
> > 
> > 	spin_lock(&dentry->d_lock);
> > 	res = dentry->d_parent->d_inode->i_ino;
> > 	spin_unlock(&dentry->d_lock);
> > 
> > Since dentry is locked and referenced, ->d_parent is stable and positive.
> > 
> > In the encrypt+casefold patch I was reviewing, it's indeed necessary, but only
> > because there was a check of inode->i_flags added outside the locked region.
> > The following would be simpler:
> > 
> >         spin_lock(&dentry->d_lock);
> >         dir = dentry->d_parent->d_inode;
> >         *pino = dir->i_ino;
> >         needs_recovery = IS_ENCRYPTED(dir) && IS_CASEFOLDED(dir);
> >         spin_unlock(&dentry->d_lock);
> > 
> > BTW, d_find_any_alias() is unnecessary too.  This code should just be using
> > file_dentry(file) from f2fs_do_sync_file().
> > 
> 
> Also, what is this code trying to accomplish?  If it's trying to find the parent
> directory of an inode with i_nlink == 1, this isn't the correct way to do it.
> The fsync could be done via a deleted file, which would make the wrong p_ino be
> set.  I think the correct approach would be to iterate through all the dentry's
> aliases, and choose the parent directory that's !IS_DEADDIR().

The intention is to give a chance to recover the pino to avoid performance
drop on fsync() by avoiding checkpoint(). And the purpose of this is to
find a file having single linked file. Otherwise, we'll do checkpoint().

> 
> - Eric