Re: [PATCH 1/4] btrfs: use BTRFS_PATH_AUTO_FREE in insert_balance_item()

[Filipe Manana]

On Tue, Apr 8, 2025 at 6:36 PM Daniel Vacek <neelx@xxxxxxxx> wrote:
>
> On Tue, 8 Apr 2025 at 16:47, Filipe Manana <fdmanana@xxxxxxxxxx> wrote:
> >
> > On Tue, Apr 8, 2025 at 1:18 PM Yangtao Li <frank.li@xxxxxxxx> wrote:
> > >
> > > All cleanup paths lead to btrfs_path_free so we can define path with the
> > > automatic free callback.
> > >
> > > Signed-off-by: Yangtao Li <frank.li@xxxxxxxx>
> > > ---
> > >  fs/btrfs/volumes.c | 7 ++-----
> > >  1 file changed, 2 insertions(+), 5 deletions(-)
> > >
> > > diff --git a/fs/btrfs/volumes.c b/fs/btrfs/volumes.c
> > > index c8c21c55be53..a962efaec4ea 100644
> > > --- a/fs/btrfs/volumes.c
> > > +++ b/fs/btrfs/volumes.c
> > > @@ -3730,7 +3730,7 @@ static int insert_balance_item(struct btrfs_fs_info *fs_info,
> > >         struct btrfs_trans_handle *trans;
> > >         struct btrfs_balance_item *item;
> > >         struct btrfs_disk_balance_args disk_bargs;
> > > -       struct btrfs_path *path;
> > > +       BTRFS_PATH_AUTO_FREE(path);
> > >         struct extent_buffer *leaf;
> > >         struct btrfs_key key;
> > >         int ret, err;
> > > @@ -3740,10 +3740,8 @@ static int insert_balance_item(struct btrfs_fs_info *fs_info,
> > >                 return -ENOMEM;
> > >
> > >         trans = btrfs_start_transaction(root, 0);
> > > -       if (IS_ERR(trans)) {
> > > -               btrfs_free_path(path);
> > > +       if (IS_ERR(trans))
> > >                 return PTR_ERR(trans);
> > > -       }
> > >
> > >         key.objectid = BTRFS_BALANCE_OBJECTID;
> > >         key.type = BTRFS_TEMPORARY_ITEM_KEY;
> > > @@ -3767,7 +3765,6 @@ static int insert_balance_item(struct btrfs_fs_info *fs_info,
> > >         btrfs_set_balance_sys(leaf, item, &disk_bargs);
> > >         btrfs_set_balance_flags(leaf, item, bctl->flags);
> > >  out:
> > > -       btrfs_free_path(path);
> > >         err = btrfs_commit_transaction(trans);
> >
> > This isn't a good idea at all.
> > We're now committing a transaction while holding a write lock on some
> > leaf of the tree root - this can result in a deadlock as the
> > transaction commit needs to update the tree root (see
> > update_cowonly_root()).
>
> I do not follow. This actually looks good to me.

path->nodes[0] has a write locked leaf, returned by btrfs_insert_empty_item().

> Is there really any functional change? What am I missing?

Yes there is, a huge one. Even if a transaction commit didn't need to
update the root tree, it would be performance wise to commit a
transaction while holding a lock on a leaf unnecessarily.

>
> --nX
>
> > Thanks.
> >
> >
> > >         if (err && !ret)
> > >                 ret = err;
> > > --
> > > 2.39.0
> > >
> > >
> >