On Fri, Oct 26, 2018 at 2:36 PM Eric Sandeen <sandeen@xxxxxxxxxxx> wrote: > On 10/26/18 12:29 AM, Stefan Ring wrote: > > On Fri, Oct 26, 2018 at 2:36 AM Eric Sandeen <sandeen@xxxxxxxxxxx> wrote: > >>> Hm, I think another case we're missing is interior nodes on multi-fsb objects, > >> i.e. everything under /* Zero unused part of interior nodes */ seems to be > >> missing for the process_multi_fsb_objects case as well. Urk, what a mess. > > > > It is a classic game of whack-a-mole ;). I do understand that it is > > this way for historical reasons, but if it would be redone today, it > > should obviously use zeroed buffers for writing out and copy just the > > bits that are known to be needed. > > Yeah, I was thinking about that last night. I'm not sure it'd be a whole > lot easier, though - for every single type of metadata block we either > need to zap the holes or copy the bits, but every boundary is the same > and would need to be computed and handled regardless. > I'm not sure it'd be a win. Yes, it would not be any easier. But it would be (amost) immediately obvious when something was missed.
