Re: [PATCH 3/3] btrfs: extended inode refs

[Chris Mason <chris.mason@oracle.com>]

On Thu, May 10, 2012 at 10:23:28AM +0200, Jan Schmidt wrote:
> On Wed, May 09, 2012 at 19:02 (+0200), Chris Mason wrote:
> >>>> +
> >>>> +	while (1) {
> >>>> +		ret = btrfs_find_one_extref(fs_root, inum, offset, path, &iref2,
> >>>> +					    &offset);
> >>>> +		if (ret < 0)
> >>>> +			break;
> >>>> +		if (ret) {
> >>>> +			ret = found ? 0 : -ENOENT;
> >>>> +			break;
> >>>> +		}
> >>>> +		++found;
> >>>> +
> >>>> +		slot = path->slots[0];
> >>>> +		eb = path->nodes[0];
> >>>> +		/* make sure we can use eb after releasing the path */
> >>>> +		atomic_inc(&eb->refs);
> >>>
> >>> You need a blocking read lock here, too. Grab it before releasing the path.
> > 
> > If you're calling btrfs_search_slot, it will give you a blocking lock
> > on the leaf.  If you set path->leave_spinning before the call, you'll
> > have a spinning lock on the leaf.
> > 
> > If you unlock a block that you got from a path (like eb =
> > path->nodes[0]), the path structure has a flag for each level that
> > indicates if that block was locked or not.  See btrfs_release_path().
> > So, don't fiddle the locks without fiddling the paths.
> > 
> > You can switch from spinning to/from blocking without touching the path, it
> > figures that out.
> 
> Note that we're releasing the path shortly after. My suggestion was to
> grab an *additional* read lock after atomic_inc(&eb->refs) (probably
> better extent_buffer_get(eb)) and before btrfs_path_release().
> 
> An alternative would be to set path->locks[0] to NULL, which would
> btrfs_release_path prevent from unlocking it, kidnapping its lock for
> our purpose. But I much prefer the open coded solution.

Thanks, I see now.

-chris