Re: [PATCH 6/7] btrfs: Don't BUG_ON alloc_path errors in find_next_chunk

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

Excerpts from Mark Fasheh's message of 2011-07-18 17:36:57 -0400:
> Hi Tsutomu,
> 
>     Thanks for the review, it is appreciated!
> 
> On Fri, Jul 15, 2011 at 11:43:52AM +0900, Tsutomu Itoh wrote:
> > > @@ -1037,7 +1037,8 @@ static noinline int find_next_chunk(struct btrfs_root *root,
> > >      struct btrfs_key found_key;
> > >  
> > >      path = btrfs_alloc_path();
> > > -    BUG_ON(!path);
> > > +    if (!path)
> > > +        return -ENOMEM;
> > 
> > If find_next_chunk() returns -ENOMEM, space_info->full becomes 1 by following code.
> > 
> > 3205 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
> > 3206                           struct btrfs_root *extent_root, u64 alloc_bytes,
> > 3207                           u64 flags, int force)
> > 3208 {
> > ...
> > 3277         ret = btrfs_alloc_chunk(trans, extent_root, flags);
> > 3278         spin_lock(&space_info->lock);
> > 3279         if (ret)
> > 3280                 space_info->full = 1;
> > 3281         else
> > 3282                 ret = 1;
> > 
> > Is it OK?
> 
> I don't think so actually. It looks like in this case we might want to
> bubble the error back up past do_chunk_alloc and leave space_info untouched.
> Chris, does that seem reasonable?

Yeah, once space_info->full is 1, we don't flip it back to zero until
more space is available somehow.  We should bubble the error up.

-chris