Re: [PATCH-v4 6/7] ext4: add support for a lazytime mount option

[Jan Kara <jack@suse.cz>]

On Thu 27-11-14 10:35:37, Dave Chinner wrote:
> On Wed, Nov 26, 2014 at 04:10:44PM -0700, Andreas Dilger wrote:
> > On Nov 26, 2014, at 3:48 PM, Dave Chinner <david@fromorbit.com> wrote:
> > > 
> > > On Wed, Nov 26, 2014 at 05:23:56AM -0500, Theodore Ts'o wrote:
> > >> Add an optimization for the MS_LAZYTIME mount option so that we will
> > >> opportunistically write out any inodes with the I_DIRTY_TIME flag set
> > >> in a particular inode table block when we need to update some inode
> > >> in that inode table block anyway.
> > >> 
> > >> Also add some temporary code so that we can set the lazytime mount
> > >> option without needing a modified /sbin/mount program which can set
> > >> MS_LAZYTIME.  We can eventually make this go away once util-linux has
> > >> added support.
> > >> 
> > >> Google-Bug-Id: 18297052
> > >> 
> > >> Signed-off-by: Theodore Ts'o <tytso@mit.edu>
> > >> ---
> > >> fs/ext4/inode.c             | 49 ++++++++++++++++++++++++++++++++++++++++++---
> > >> fs/ext4/super.c             |  9 +++++++++
> > >> include/trace/events/ext4.h | 30 +++++++++++++++++++++++++++
> > >> 3 files changed, 85 insertions(+), 3 deletions(-)
> > >> 
> > >> diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c
> > >> index 5653fa4..8308c82 100644
> > >> --- a/fs/ext4/inode.c
> > >> +++ b/fs/ext4/inode.c
> > >> @@ -4140,6 +4140,51 @@ static int ext4_inode_blocks_set(handle_t *handle,
> > >> }
> > >> 
> > >> /*
> > >> + * Opportunistically update the other time fields for other inodes in
> > >> + * the same inode table block.
> > >> + */
> > >> +static void ext4_update_other_inodes_time(struct super_block *sb,
> > >> +					  unsigned long orig_ino, char *buf)
> > >> +{
> > >> +	struct ext4_inode_info	*ei;
> > >> +	struct ext4_inode	*raw_inode;
> > >> +	unsigned long		ino;
> > >> +	struct inode		*inode;
> > >> +	int		i, inodes_per_block = EXT4_SB(sb)->s_inodes_per_block;
> > >> +	int		inode_size = EXT4_INODE_SIZE(sb);
> > >> +
> > >> +	ino = orig_ino & ~(inodes_per_block - 1);
> > >> +	for (i = 0; i < inodes_per_block; i++, ino++, buf += inode_size) {
> > >> +		if (ino == orig_ino)
> > >> +			continue;
> > >> +		inode = find_active_inode_nowait(sb, ino);
> > >> +		if (!inode ||
> > >> +		    (inode->i_state & I_DIRTY_TIME) == 0 ||
> > >> +		    !spin_trylock(&inode->i_lock)) {
> > >> +			iput(inode);
> > >> +			continue;
> > >> +		}
> > >> +		inode->i_state &= ~I_DIRTY_TIME;
> > >> +		inode->i_ts_dirty_day = 0;
> > >> +		spin_unlock(&inode->i_lock);
> > >> +		inode_requeue_dirtytime(inode);
> > >> +
> > >> +		ei = EXT4_I(inode);
> > >> +		raw_inode = (struct ext4_inode *) buf;
> > >> +
> > >> +		spin_lock(&ei->i_raw_lock);
> > >> +		EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
> > >> +		EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
> > >> +		EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
> > >> +		ext4_inode_csum_set(inode, raw_inode, ei);
> > >> +		spin_unlock(&ei->i_raw_lock);
> > >> +		trace_ext4_other_inode_update_time(inode, orig_ino);
> > >> +		iput(inode);
> > >> +	}
> > >> +}
> > > 
> > > Am I right in that this now does unlogged timestamp updates of
> > > inodes? What happens when that buffer gets overwritten by log
> > > recover after a crash? The timestamp updates get lost?
> > > 
> > > FYI, XFS has had all sorts of nasty log recovery corner cases
> > > caused by log recovery overwriting non-logged inode updates like
> > > this. In the past few years we've removed every single non-logged
> > > inode update "optimisation" so that all changes (including timestamps)
> > > are transactional so inode state on disk not matching what log
> > > recovery wrote to disk for all the other inode metadata...
> > > 
> > > Optimistic unlogged inode updates are a slippery slope, and history
> > > tells me that it doesn't lead to a nice place....
> > 
> > Since ext4/jbd2 is logging the whole block, unlike XFS which is doing
> > logical journaling, this isn't an unlogged update.  It is just taking
> > advantage of the fact that the whole block is going to be logged and
> > written to the disk anyway.
> 
> Urk - that's worse, isn't it? i.e the code above calls iput() from
> within a current transaction context?  What happens if that drops
> the last reference to the inode and it gets evicted due to racing
> with an unlink? Won't that try to start another transaction to free
> the inode (i.e. through ext4_evict_inode())?
  Yeah, the patch looks buggy (and racy wrt concurrent updates of time
stamps as well). I think if we want to do this optimization, we would need
a function like "clear inode dirty bits for this range of inode numbers".
That is doable atomically within VFS and although it looks somewhat ugly,
the performance
-- 
Jan Kara <jack@suse.cz>
SUSE Labs, CR