On Mon, Oct 21, 2019 at 10:29:58AM +1100, Dave Chinner wrote: > This is, IMO, a step backards. We're going to end up failing to > initialise new fields correctly with this... How is that different from the plain xfs_inode fields? In fact I suspect most of the initializers cn just be removed entirely, so I'll look into preloading that at the front of the series. > This is a bug and should make all the 32-bit project ID tests fail. > If it doesn't them we've got a problem with our test coverage. If it > does fail, then I'm not sure this patchset has been adequately > tested... I don't think we have any coverage of that, at least I didn't see any extra failures. > > + xfs_fsize_t i_disk_size; /* number of bytes in file */ > > + xfs_rfsblock_t i_nblocks; /* direct & btree blocks used */ > > + xfs_extlen_t i_extsize; /* extent size hint */ > > + xfs_extnum_t i_nextents; /* # of extents in data fork */ > > + xfs_aextnum_t i_anextents; /* # of extents in attr fork */ > > + uint8_t i_forkoff; /* attr fork offset */ > > + int8_t i_aformat; /* attr fork format */ > > + uint32_t i_dmevmask; /* DMIG event mask */ > > + uint16_t i_dmstate; /* DMIG state info */ > > If we are cleaning up the icdinode, why do these still exist in > memory? Because we need them so that we put the right value in the log when logging the inode core. Otherwise a log recovery might clear these values. The only thing I could do is add a log incompat flag set on a kernel that removes the field and then not apply changes to these two fields when recoverying the log on a file system with that flag set.