On Thu, Nov 27, 2014 at 04:41:59PM +0100, Jan Kara wrote:

  Hum, but this puts lots of stuff under inode_hash_lock, including
writeback list lock. I don't like this too much. I understand that getting
handle for each inode is rather more CPU intensive but it should still be a
clear win over the current situation and avoids entangling locks like this.

Hmm, if we dropped the inode_requeue_dirtytime(), then we can avoid
taking the writeback list lock.  The net result is that we would have
some inodes still on the b_dirty_time list that were no longer
I_DIRTY_TIME, but since I_DIRTY_TIME wouldn't be set, it's mostly
harmless since when we do iterate over the b_dirty_time list, those
inodes can be quickly identified and skipped over.  (And if the inode
ever gets dirtied for real, then it will get moved onto the b_dirty
list and that will be that.)

The problem with getting a handle on the inode is not just that it is
more CPU intensive, but that can't let the iput_final() call happen
until after we have finished the transaction handle.  We could keep a
linked list of inodes attached to the handle, and then only call iput
on them once ext4_journal_stop(handle) gets called, but that's a
complication I'd like to avoid if at all possible.

Being able to opporunistically write the timestamps when we are
journalling an inode table block is a pretty big win, so if we end up
extending the hold time on inode_hash_lock (only when we come across a
I_DIRTY_TIME inode that we can clear) a tiny bit, there will be a lot
of workloads where I think it's a worthwhile tradeoff.  If we can
avoid entangling the writebakc list lock, does that make you happier
about this approach?

	     	      	       	     	 - Ted

_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs