On Thu, Oct 11, 2018 at 06:43:57PM -0600, Andreas Dilger wrote:
> I was looking at POSIX ACL on-disk and in-memory code and it looks like
> there is a subtle dependency between the on-disk format and what (IMHO)
> would be considered in-memory declarations.
>
> When a POSIX ACL is read from disk, posix_acl_from_mode() copies the file
> mode (S_I[RWX][UGO]) into the e_perm fields of the ACL default entries.
> Similarly, in posix_acl_equiv_mode() and posix_acl_create_masq() it uses
> S_IRWXO to mask the e_perm flags.
>
> However, later on in posix_acl_permission() it directly uses the "want"
> flag contains MAY_{READ,WRITE,EXEC} flags and compares those to e_perm of
> each ACL entry.
>
> In posix_acl_valid() it compares e_perm with ACL_{READ,WRITE,EXECUTE}.
>
> While the MAY_[RWX] and ACL_[RWX] currently have the same value as
> S_I[RWX]OTH, it isn't very clear that these flags MUST all have the same
> values or POSIX ACLs will break.
>
> This definitely doesn't seem quite right. Are the ACL_* constants the
> values to be used, with "conversion" in between the flags/modes? Should
> there be a BUILD_BUG_ON() that trips if those values ever differ?
Encoding of rwx bits is pretty much cast in stone - they go all way back
to v1 (if not to PDP7 times) and I can't imagine any Unix variant that
would have them not in the same order. MAY_... is tied to those and
so are the bits in ->e_perm.
IOW, all of those are in practice immutable - too closely tied to on-disk
data structures in a lot of filesystems *and* to any number of userland
programs using explicit octal values for mkdir(), etc. arguments.
Symbolic constants != can realistically be redefined...
