Re: [PATCH 15/26] Btrfs: add a new source file with device replace code

[Liu Bo <bo.li.liu@oracle.com>]

On Thu, Nov 08, 2012 at 06:24:36PM +0100, Stefan Behrens wrote:
> On Thu, 8 Nov 2012 22:50:47 +0800, Liu Bo wrote:
> > On Tue, Nov 06, 2012 at 05:38:33PM +0100, Stefan Behrens wrote:
[...]
> >> +	btrfs_dev_replace_unlock(dev_replace);
> >> +
> >> +	btrfs_wait_ordered_extents(root, 0);
> >> +
> >> +	/* force writing the updated state information to disk */
> >> +	trans = btrfs_start_transaction(root, 0);
> > 
> > why a start_transaction here?  Any reasons?
> > (same question also for some other places)
> > 
> 
> Without this transaction, there is outstanding I/O which is not flushed.
> Pending writes that go only to the old disk need to be flushed before
> the mode is switched to write all live data to the source disk and to
> the target disk as well. The copy operation that is part of the scrub
> code works on the commit root for performance reasons. Every write
> request that is performed after the commit root is established needs to
> go to both disks. Those requests that already have the bdev assigned
> (i.e., btrfs_map_bio() was already called) cannot be duplicated anymore
> to write to the new disk as well.
> 
> btrfs_dev_replace_finishing() looks similar and goes through a
> transaction commit between the steps where the bdev in the mapping tree
> is swapped and the step when the old bdev is freed. Otherwise the bdev
> would be accessed after being freed.
> 

I see, if you're only about to flush metadata, why not join a transaction?

thanks,
liubo