Re: RFC: log record CRC validation

[Andi Kleen <andi@xxxxxxxxxxxxxx>]

David Chinner <dgc@xxxxxxx> writes:
 
> I haven't looked at great depth into other structures in terms of
> implementation details. I know that if we use a 16 bit CRC on
> directories we can get away without a on-disk format change as the
> xfs_da_blkinfo structure has 16 bits of padding. However, given that
> directory block size can reach 64k, a CRC16 check is really only
> capable of single bit error detection. Hence I think we really need
> CRC32 here which means an on-disk format change.

When the directory format is changed it would be nice to also support
DT_* types at the same time. They can speed up some operations nicely
because file system walkers can avoid a stat() (and seek to the inode)
just to find out if a name is a directory or not. Right now there is
no space for this unfortunately.

-Andi