On Wed, Dec 18, 2019 at 01:43:09PM -0800, Darrick J. Wong wrote: > > diff --git a/fs/xfs/libxfs/xfs_attr.h b/fs/xfs/libxfs/xfs_attr.h > > index 91c2cb14276e..04a628016728 100644 > > --- a/fs/xfs/libxfs/xfs_attr.h > > +++ b/fs/xfs/libxfs/xfs_attr.h > > @@ -36,11 +36,10 @@ struct xfs_attr_list_context; > > #define ATTR_KERNOTIME 0x1000 /* [kernel] don't update inode timestamps */ > > #define ATTR_KERNOVAL 0x2000 /* [kernel] get attr size only, not value */ > > > > -#define ATTR_INCOMPLETE 0x4000 /* [kernel] return INCOMPLETE attr keys */ > > Hmm, did we used to allow ATTR_INCOMPLETE from the attr_multi userspace > calls? We allowed that, but it didn't do anything as only the attr list code checked for ATTR_INCOMPLETE. > Maybe we should leave ATTR_INCOMPLETE so that we can blacklist > it in case we ever see it coming from userspace? Or at least prevent > anyone from reusing 0x4000 and suffering the confusion. At then end of the series the whole flags mess is cleaned up and there is a hard barrier from user to kernel flags, so this should all be set. > I also wonder if we can break userspace this way, but OTOH userspace > should never be able to query incomplete attrs and this is an obscure > ioctl anyway so maybe it's fine.... Exactly. Incomplete attrs are an implementation detail that must not leak to userspace.
