Re: [PATCH] Btrfs: do not abort transaction when failing to insert hole extent

[David Sterba <dsterba@suse.cz>]

On Thu, Apr 05, 2018 at 11:58:16AM -0700, Liu Bo wrote:
> On Thu, Apr 5, 2018 at 9:48 AM, David Sterba <dsterba@suse.cz> wrote:
> > On Sat, Mar 31, 2018 at 06:11:55AM +0800, Liu Bo wrote:
> >> This is running in a typical write path, not inside a critical path
> >> where we have to abort the running transaction, so it's OK to return
> >> errors to callers and eventually to userspace.
> >
> > I'm not sure this is entierly correct, several other places do not abort
> > after btrfs_drop_extents as there's nothing that would leave the
> > structres in some half-state.
> >
> >> Signed-off-by: Liu Bo <bo.liu@linux.alibaba.com>
> >> ---
> >>  fs/btrfs/inode.c | 5 +----
> >>  1 file changed, 1 insertion(+), 4 deletions(-)
> >>
> >> diff --git a/fs/btrfs/inode.c b/fs/btrfs/inode.c
> >> index c7b75dd..b9310f8 100644
> >> --- a/fs/btrfs/inode.c
> >> +++ b/fs/btrfs/inode.c
> >> @@ -4939,16 +4939,13 @@ static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
> >>
> >>       ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
> >>       if (ret) {
> >> -             btrfs_abort_transaction(trans, ret);
> >>               btrfs_end_transaction(trans);
> >>               return ret;
> >>       }
> >>
> >>       ret = btrfs_insert_file_extent(trans, root, btrfs_ino(BTRFS_I(inode)),
> >>                       offset, 0, 0, len, 0, len, 0, 0, 0);
> >
> > But here the extents have been already dropped and missing to insert the
> > items does not seem to lead to a consistent state.
> >
> > It's possible that I'm missing something. In a call path that can be
> > safely rolled back even with a started transaction, we don't need to
> > abort in all cases. But if the rollback requires some non-trivial
> > modifications, I don't see options how to avoid the abort.
> >
> > __btrfs_drop_extents does a lot of state changes and can itself fail
> > in the middle of dropping the range, aborting looks like the safest
> > option.
> >
> 
> As maybe_insert_hole is only called by btrfs_cont_expand here, which
> means it's a really hole, I don't expect drop_extents would drop
> anything, we can remove this drop_extents and put an assert after
> btrfs_insert_file_extent for checking EEXIST.

Sounds good.

> It's different from punch hole where we need to explicitly drop an
> actual extent and replace it with a hole range.

Right, that's what I didn't see at first.