Re: [RFC v7 21/41] richacl: Move everyone@ aces down the acl

On Mon, Sep 21, 2015 at 11:43:16PM +0200, Andreas Gruenbacher wrote:

2015-09-18 21:35 GMT+02:00 J. Bruce Fields [off-list ref]:

quoted

On Sat, Sep 05, 2015 at 12:27:16PM +0200, Andreas Gruenbacher wrote:

quoted

The POSIX standard puts processes which are not the owner or a member in
the owning group or which match any ace other then everyone@ on the
other file class.  We only know if a process is in the other class after
processing the entire acl.

Move all everyone@ aces in the acl down in the acl so that at most a
single everyone@ allow ace remains at the end.  Permissions which are
not explicitly allowed are implicitly denied, so an everyone@ deny ace
is unneeded.

The everyone@ aces can be moved down the acl without changing the
permissions that the acl grants.  This transformation simplifies the
following algorithms, and eventually allows us to turn the final
everyone@ allow ace into an entry for the other class.

Signed-off-by: Andreas Gruenbacher <redacted>
---
 fs/richacl_compat.c | 65 +++++++++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 65 insertions(+)

diff --git a/fs/richacl_compat.c b/fs/richacl_compat.c
index 341e429..4f0acf5 100644
--- a/fs/richacl_compat.c
+++ b/fs/richacl_compat.c

@@ -153,3 +153,68 @@ richace_change_mask(struct richacl_alloc *alloc, struct richace **ace,
      }
      return 0;
 }
+
+/**
+ * richacl_move_everyone_aces_down  -  move everyone@ aces to the end of the acl
+ * @alloc:   acl and number of allocated entries
+ *
+ * Move all everyone aces to the end of the acl so that only a single everyone@
+ * allow ace remains at the end, and update the mask fields of all aces on the
+ * way.  The last ace of the resulting acl will be an everyone@ allow ace only
+ * if @acl grants any permissions to @everyone.  No @everyone deny aces will
+ * remain.
+ *
+ * This transformation does not alter the permissions that the acl grants.
+ * Having at most one everyone@ allow ace at the end of the acl helps us in the
+ * following algorithms.
+ */
+static int
+richacl_move_everyone_aces_down(struct richacl_alloc *alloc)
+{
+     struct richace *ace;
+     unsigned int allowed = 0, denied = 0;
+
+     richacl_for_each_entry(ace, alloc->acl) {
+             if (richace_is_inherit_only(ace))
+                     continue;
+             if (richace_is_everyone(ace)) {
+                     if (richace_is_allow(ace))
+                             allowed |= (ace->e_mask & ~denied);
+                     else if (richace_is_deny(ace))
+                             denied |= (ace->e_mask & ~allowed);
+                     else
+                             continue;
+                     if (richace_change_mask(alloc, &ace, 0))
+                             return -1;
+             } else {
+                     if (richace_is_allow(ace)) {
+                             if (richace_change_mask(alloc, &ace, allowed |
+                                             (ace->e_mask & ~denied)))
+                                     return -1;
+                     } else if (richace_is_deny(ace)) {
+                             if (richace_change_mask(alloc, &ace, denied |
+                                             (ace->e_mask & ~allowed)))
+                                     return -1;
+                     }
+             }
+     }
+     if (allowed & ~RICHACE_POSIX_ALWAYS_ALLOWED) {
+             struct richace *last_ace = ace - 1;
+
+             if (alloc->acl->a_entries &&
+                 richace_is_everyone(last_ace) &&
+                 richace_is_allow(last_ace) &&
+                 richace_is_inherit_only(last_ace) &&
+                 last_ace->e_mask == allowed)
+                     last_ace->e_flags &= ~RICHACE_INHERIT_ONLY_ACE;

That's a funny special case!  Is it even worth it, or could we just live
with an extra uninheritable EVERYONE ace in this case?

Inheritable everyone@ allow entries at the end of the ACL are not
uncommon. This special case prevents the algorithm from splitting such
entries into inherit-only and non-inheritable parts.

OK.--b.