Re: [PATCH] skip I_CLEAR state inodes

[Wu Fengguang <fengguang.wu@intel.com>]

On Tue, Jun 02, 2009 at 07:37:36PM +0800, Jan Kara wrote:
> On Tue 02-06-09 16:55:23, Wu Fengguang wrote:
> > On Tue, Jun 02, 2009 at 05:38:35AM +0800, Eric Sandeen wrote:
> > > Wu Fengguang wrote:
> > > > Add I_CLEAR tests to drop_pagecache_sb(), generic_sync_sb_inodes() and
> > > > add_dquot_ref().
> > > > 
> > > > clear_inode() will switch inode state from I_FREEING to I_CLEAR,
> > > > and do so _outside_ of inode_lock. So any I_FREEING testing is
> > > > incomplete without the testing of I_CLEAR.
> > > > 
> > > > Masayoshi MIZUMA first discovered the bug in drop_pagecache_sb() and
> > > > Jan Kara reminds fixing the other two cases. Thanks!
> > > 
> > > Is there a reason it's not done for __sync_single_inode as well?
> > 
> > It missed the glance because it don't have an obvious '|' in the line ;)
> > 
> > > Jeff Layton asked the question and I'm following it up :)
> > > 
> > > __sync_single_inode currently only tests I_FREEING, but I think we are
> > > safe because __sync_single_inode sets I_SYNC, and clear_inode waits for
> > > I_SYNC to be cleared before it changes I_STATE.
> > 
> > But I_SYNC is removed just before the I_FREEING test, so we still have
> > a small race window?
> > 
> > > On the other hand, testing I_CLEAR here probably would be safe anyway,
> > > and it'd be bonus points for consistency?
> > 
> > So let's add the I_CLEAR test?
> > 
> > > Same basic question for generic_sync_sb_inodes, which has a
> > > BUG_ON(inode->i_state & I_FREEING), seems like this could check I_CLWAR
> > > as well?
> > 
> > Yes, we can add I_CLEAR here to catch more error condition.
> > 
> > Thanks,
> > Fengguang
> > 
> > ---
> > skip I_CLEAR state inodes in writeback routines
> > 
> > The I_FREEING test in __sync_single_inode() is racy because
> > clear_inode() can set i_state to I_CLEAR between the clear of I_SYNC
> > and the test of I_FREEING.
> > 
> > Also extend the coverage of BUG_ON(I_FREEING) to I_CLEAR.
> > 
> > Reported-by: Jeff Layton <jlayton@redhat.com>
> > Reported-by: Eric Sandeen <sandeen@sandeen.net>
> > Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
> > ---
> >  fs/fs-writeback.c |    4 ++--
> >  1 file changed, 2 insertions(+), 2 deletions(-)
> > 
> > --- linux.orig/fs/fs-writeback.c
> > +++ linux/fs/fs-writeback.c
> > @@ -316,7 +316,7 @@ __sync_single_inode(struct inode *inode,
> >  	spin_lock(&inode_lock);
> >  	WARN_ON(inode->i_state & I_NEW);
> >  	inode->i_state &= ~I_SYNC;
> > -	if (!(inode->i_state & I_FREEING)) {
> > +	if (!(inode->i_state & (I_FREEING | I_CLEAR))) {
> >  		if (!(inode->i_state & I_DIRTY) &&
> >  		    mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
>   Is the whole if needed? I had an impression that everyone calling
> __sync_single_inode() should better take care it does not race with inode
> freeing... So WARN_ON would be more appropriate IMHO.
> 
> >  			/*
> > @@ -518,7 +518,7 @@ void generic_sync_sb_inodes(struct super
> >  		if (current_is_pdflush() && !writeback_acquire(bdi))
> >  			break;
> >  
> > -		BUG_ON(inode->i_state & I_FREEING);
> > +		BUG_ON(inode->i_state & (I_FREEING | I_CLEAR));
> >  		__iget(inode);
> >  		pages_skipped = wbc->pages_skipped;
> >  		__writeback_single_inode(inode, wbc);
>   Looking at this code, it looks a bit suspicious. What prevents this s_io
> list scan to race with inode freeing? In particular generic_forget_inode()

Good catch.

> can drop inode_lock to write the inode and in the mean time
> generic_sync_sb_inodes() can come, get a reference to the inode and start
> it's writeback... Subsequent iput() would then call generic_forget_inode()

Another possibility:

generic_forget_inode
  inode->i_state |= I_WILL_FREE;
  spin_unlock(&inode_lock);
                                                generic_sync_sb_inodes()
                                                  spin_lock(&inode_lock);
                                                  __iget(inode);
                                                  __writeback_single_inode
                                                    // see non zero i_count
                                                    WARN_ON(inode->i_state & I_WILL_FREE);

I'm wondering why didn't we saw reports on the last WARN_ON()?
Did we missed something?

> on the inode again. So shouldn't we skip I_FREEING|I_CLEAR|I_WILL_FREE|I_NEW
> inodes in this scan like we do for later in the function for another scan?