Re: generic_permission() optimization

[Al Viro <viro@zeniv.linux.org.uk>]

On Thu, Oct 31, 2024 at 08:14:59AM -1000, Linus Torvalds wrote:
> On Wed, 30 Oct 2024 at 20:42, Linus Torvalds
> <torvalds@linux-foundation.org> wrote:
> >
> > I was kind of hoping that such cases would use 'cache_no_acl()' which
> > makes that inode->i_acl be NULL. Wouldn't that be the right model
> > anyway for !IS_POSIXACL()?
> 
> Alternatively, just initialize it to NULL in inode_init_always_gfp(), eg like
> 
> -       inode->i_acl = inode->i_default_acl = ACL_NOT_CACHED;
> +       inode->i_acl = inode->i_default_acl =
> +               (sb->s_flags & SB_POSIXACL) ? ACL_NOT_CACHED : NULL;

IIRC, the reason we do not do that was the possibility of mount -o remount,acl
Not that it makes much sense, but it's currently supported and POSIX ACLs
don't make much sense to start with...

Anyway, patch looks sane; I still think that adding || !IS_POSIXACL(inode)
wouldn't hurt (yes, it's a dereference of ->i_sb in case when ->i_acl
is ACL_NOT_CACHED, but we are going to dereference it shortly after
in case we don't take the fast path.  OTOH, that probably matters only
for fairly specific loads - massive accesses to procfs and sysfs, mostly.