David Sterba <dsterba@xxxxxxx> wrote: > The ordering of the bits set/cleared can prevent unwanted combinations, > like: when the COMPRESS bit is set, there's no NODATACOW anymore: > > - clear NODATACOW > - set compress_type > - set COMPRESS > > I haven't checked with the sources if this is so. The current code doesn't use memory barriers or atomic ops, so you must assume that those memory operations can be reordered/merged by the compiler or the cpu. > So the changes would be possible in one atomic (bitwise) call? Instead > of: > > clear_bit(opts, NODATACOW) > set_bit(opts, COMPRESS) > > do: > > set_bits_with_mask(opts, COMPRESS, BTRFS_DATA__STORAGE_OPT) After the mount API changes, the value for the variable will be composed in a separate context and then applied with a single assignment after printing any messages about changes in active options. > Side note: up to now we're all used to the NO* naming and semantics of > the feature bits, so it might take some time to do the mental switch. Yeah - I've actually switched them all to positive naming in what I have so far as it simplifies the code and makes it easier to think about (for me at least). David
