Re: [PATCH] fs / ext3: Always unlock updates in ext3_freeze()

[Jan Kara <jack@suse.cz>]

  Hello,

On Mon 15-08-11 20:09:13, Rafael J. Wysocki wrote:
> On Monday, August 15, 2011, Jan Kara wrote:
> >   Hi,
> > On Thu 11-08-11 23:31:13, Rafael J. Wysocki wrote:
> > > From: Rafael J. Wysocki <rjw@sisk.pl>
> > > 
> > > In analogy with ext4 make ext3_freeze() always call
> > > journal_unlock_updates() to prevent it from leaving a locked mutex
> > > behind.  Accordingly, modify ext3_unfreeze() so that it doesn't
> > > call journal_unlock_updates() any more.
> > > 
> > > Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
> > > ---
> > > 
> > > Sorry for the duplicate, the previous one was sent too early.
> > > 
> > > ---
> > >  fs/ext3/super.c |   39 ++++++++++++++++++---------------------
> > >  1 file changed, 18 insertions(+), 21 deletions(-)
> > > 
> > > Index: linux/fs/ext3/super.c
> > > ===================================================================
> > > --- linux.orig/fs/ext3/super.c
> > > +++ linux/fs/ext3/super.c
> > > @@ -2535,30 +2535,28 @@ static int ext3_sync_fs(struct super_blo
> > >   */
> > >  static int ext3_freeze(struct super_block *sb)
> > >  {
> > > -	int error = 0;
> > > +	int error;
> > >  	journal_t *journal;
> > >  
> > > -	if (!(sb->s_flags & MS_RDONLY)) {
> > > -		journal = EXT3_SB(sb)->s_journal;
> > > +	if (sb->s_flags & MS_RDONLY)
> > > +		return 0;
> > >  
> > > -		/* Now we set up the journal barrier. */
> > > -		journal_lock_updates(journal);
> > > +	journal = EXT3_SB(sb)->s_journal;
> > >  
> > > -		/*
> > > -		 * We don't want to clear needs_recovery flag when we failed
> > > -		 * to flush the journal.
> > > -		 */
> > > -		error = journal_flush(journal);
> > > -		if (error < 0)
> > > -			goto out;
> > > -
> > > -		/* Journal blocked and flushed, clear needs_recovery flag. */
> > > -		EXT3_CLEAR_INCOMPAT_FEATURE(sb, EXT3_FEATURE_INCOMPAT_RECOVER);
> > > -		error = ext3_commit_super(sb, EXT3_SB(sb)->s_es, 1);
> > > -		if (error)
> > > -			goto out;
> > > -	}
> > > -	return 0;
> > > +	/* Now we set up the journal barrier. */
> > > +	journal_lock_updates(journal);
> > > +
> > > +	/*
> > > +	 * We don't want to clear needs_recovery flag when we failed
> > > +	 * to flush the journal.
> > > +	 */
> > > +	error = journal_flush(journal);
> > > +	if (error < 0)
> > > +		goto out;
> > > +
> > > +	/* Journal blocked and flushed, clear needs_recovery flag. */
> > > +	EXT3_CLEAR_INCOMPAT_FEATURE(sb, EXT3_FEATURE_INCOMPAT_RECOVER);
> > > +	error = ext3_commit_super(sb, EXT3_SB(sb)->s_es, 1);
> > >  
> > >  out:
> > >  	journal_unlock_updates(journal);
> > > @@ -2577,7 +2575,6 @@ static int ext3_unfreeze(struct super_bl
> > >  		EXT3_SET_INCOMPAT_FEATURE(sb, EXT3_FEATURE_INCOMPAT_RECOVER);
> > >  		ext3_commit_super(sb, EXT3_SB(sb)->s_es, 1);
> > >  		unlock_super(sb);
> > > -		journal_unlock_updates(EXT3_SB(sb)->s_journal);
> > >  	}
> > >  	return 0;
> > >  }
> >   It's not so simple as this. Ext3 relies on the mutex (the one hidden in
> > journal_lock_updates()) to make sure that new transaction cannot be started
> > while the filesystem is frozen - that's essentially what makes the
> > filesystem frozen. So if we want to get rid of the mutex we have to achieve
> > blocking by something else - ext4 uses vfs_check_frozen() in
> > ext4_journal_start().
> 
> I see.  Still, freeze_bdev() may be called by user space through a syscall,
> as far as I can say, so it shouldn't leave the mutex locked.
  Yes, I agree with you. That's an ugliness left over from a long time ago.
I'll have a look at fixing this...

> >   BTW,  filesystem freezing never really worked for mmaped writes under
> > ext3 - ext3 would have to implement page_mkwrite() callback for that - so
> > if you want to rely on it for suspending, this will be non-trivial.
> 
> At this point the purpose of freezing filesystems is basically to
> prevent XFS from deadlocking with hibernation's memory preallocation.
> For other filesystems it may or may not make a difference depending on
> their implementation of freeze/unfreeze_super().
  What's exactly the problem? Memory preallocation enters direct reclaim
and that deadlocks in the filesystem?

								Honza