Date:	Mon, 14 Sep 2009 17:25:29 +0200
From:	Jan Kara <jack@...e.cz>
To:	Mingming <cmm@...ibm.com>
Cc:	Jan Kara <jack@...e.cz>, Theodore Tso <tytso@....edu>,
	"Aneesh Kumar K.V" <aneesh.kumar@...ux.vnet.ibm.com>,
	linux-ext4@...r.kernel.org
Subject: Re: fsync on ext[34] working only by an accident

On Thu 10-09-09 13:14:46, Mingming wrote:
> On Thu, 2009-09-10 at 16:06 +0200, Jan Kara wrote:
> > On Thu 10-09-09 09:10:07, Theodore Tso wrote:
> > > On Thu, Sep 10, 2009 at 04:34:55PM +0530, Aneesh Kumar K.V wrote:
> > > > mark_buffer_dirty -> __set_page_dirty -> __mark_inode_dirty
> > > 
> > > We need to be careful here.  First of all, mark_buffer_dirty() on the
> > > code paths you are talking about is being passed a metadata buffer
> > > head.  As such, has Jan has pointed out, the bh is part of the buffer
> > > cache, so the page->mapping of associated with bh->b_page is the inode
> > > of the block device --- *not* the ext4 inode.
> > > 
> > > Secondly, __set_page_dirty calls __mark_inode_dirty passing in
> > > I_DIRTY_PAGES --- which should be a hint.  What Jan is talking about
> > > is where we set the inode flags I_DIRTY_SYNC and I_DIRTY_DATASYNC:
> > > 
> > >  * I_DIRTY_SYNC		Inode is dirty, but doesn't have to be written on
> > >  *			fdatasync().  i_atime is the usual cause.
> > >  * I_DIRTY_DATASYNC	Data-related inode changes pending. We keep track of
> > >  *			these changes separately from I_DIRTY_SYNC so that we
> > >  *			don't have to write inode on fdatasync() when only
> > >  *			mtime has changed in it.
> > > 
> > > This is important because ext4_sync_file() (which is called by fsync()
> > > and fdatasync()) uses this logic to determine whether or not to call
> > > sync_inode(), which is what will force a commit when wbc.sync_mode is
> > > set to WB_SYNC_ALL.
> >   Yes, this is exactly what I was trying to point out.
> > 
> > > In fact, I think the problem is worse than Jan is pointing out,
> > > because it's not enough that vfs_fq_alloc_space() is calling
> > > mark_inode_dirty(), since that only sets I_DIRTY_SYNC.  When we touch
> > > i_size or i_block[], we need to make sure that I_DIRTY_DATASYNC is
> > > set, so that fdatasync() will force a commit.
> >   Actually no. mark_inode_dirty() dirties inode with I_DIRTY ==
> > (I_DIRTY_SYNC | I_DIRTY_DATASYNC | I_DIRTY_PAGES) so it happens to work.
> > The fact that quota *could* dirty the inode with I_DIRTY_SYNC only
> > is right but that's a separate issue.
> > 
> > > I think the right thing to do is to create an
> > > _ext[34]_mark_inode_dirty() which takes an extra argument, which
> > > controls whether or not we set I_DIRTY_SYNC or I_DIRTY_DATASYNC.  In
> > > fact, most of the time, we want to be setting I_DIRTY_DATASYNC, so we
> > > should probably have ext[34]_mark_inode_dirty() call
> > > _ext[34]_mark_inode_dirty() with I_DIRTY_DATASYNC, and then create a
> > > ext[34]_mark_inode_nodatasync() version passes in I_DIRTY_SYNC.
> > > 
> > > This will cause pdflush to call ext4_write_inode() more frequently,
> > > but pdflush calls write_inode with wait=0, and ext4_write_inode() is a
> > > no-op in that case.
> >   Thinking about it, it won't work so easily. The problem is that when
> > pdflush decides to write the inode, it unconditionally clears dirty flags.
> > We could redirty the inode in write_inode() but that's IMHO too ugly to
> > bear it.
> 
> I am a little confused here, so pdflush could found the dirty inodes
> (due to quota) but it doesn't force journal comminit and write the inode
> to disk? 
  pdflush never forces a journal commit. It just does a periodic writeout
so there's no need to force a journal commit. It just calls write_inode(),
which ext3/4 implement as NOP because the inode is already part of the
running transaction when it is dirty.

									Honza
-- 
Jan Kara <jack@...e.cz>
SUSE Labs, CR
--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Powered by blists - more mailing lists