Linux optimize access checks when ACL is trivial

Bypass check of ZFS aces if the ACL is trivial. When an ACL is
trivial its permissions are represented by the mode without any
loss of information. In this case, it is safe to convert the
access request into equivalent mode and then pass desired mask
and inode to generic_permission(). This has the added benefit
of also checking whether entries in a POSIX ACL on the file grant
the desired access.

This commit also skips the ACL check on looking up the xattr dir
since such restrictions don't exist in Linux kernel and it makes
xattr lookup behavior inconsistent between SA and file-based
xattrs. We also don't want to perform a POSIX ACL check while
looking up the POSIX ACL if for some reason it is located in
the xattr dir rather than an SA.

Reviewed-by: Brian Behlendorf <[email protected]>
Co-authored-by: Ryan Moeller <[email protected]>
Signed-off-by: Andrew Walker <[email protected]>
Closes #13237 

1 files changed, 70 insertions, 0 deletions

diff --git a/module/os/linux/zfs/zfs_acl.c b/module/os/linux/zfs/zfs_acl.c
index 351e4dad7..b70691ab3 100644
--- a/module/os/linux/zfs/zfs_acl.c
+++ b/module/os/linux/zfs/zfs_acl.c
@@ -863,6 +863,26 @@ zfs_unix_to_v4(uint32_t access_mask)
 	return (new_mask);
 }
 
+
+static int
+zfs_v4_to_unix(uint32_t access_mask, int *unmapped)
+{
+	int new_mask = 0;
+
+	*unmapped = access_mask &
+	    (ACE_WRITE_OWNER | ACE_WRITE_ACL | ACE_DELETE);
+
+	if (access_mask & WRITE_MASK)
+		new_mask |= S_IWOTH;
+	if (access_mask & ACE_READ_DATA)
+		new_mask |= S_IROTH;
+	if (access_mask & ACE_EXECUTE)
+		new_mask |= S_IXOTH;
+
+	return (new_mask);
+}
+
+
 static void
 zfs_set_ace(zfs_acl_t *aclp, void *acep, uint32_t access_mask,
     uint16_t access_type, uint64_t fuid, uint16_t entry_type)
@@ -2399,6 +2419,53 @@ zfs_has_access(znode_t *zp, cred_t *cr)
 	return (B_TRUE);
 }
 
+/*
+ * Simplified access check for case where ACL is known to not contain
+ * information beyond what is defined in the mode. In this case, we
+ * can pass along to the kernel / vfs generic_permission() check, which
+ * evaluates the mode and POSIX ACL.
+ *
+ * NFSv4 ACLs allow granting permissions that are usually relegated only
+ * to the file owner or superuser. Examples are ACE_WRITE_OWNER (chown),
+ * ACE_WRITE_ACL(chmod), and ACE_DELETE. ACE_DELETE requests must fail
+ * because with conventional posix permissions, right to delete file
+ * is determined by write bit on the parent dir.
+ *
+ * If unmappable perms are requested, then we must return EPERM
+ * and include those bits in the working_mode so that the caller of
+ * zfs_zaccess_common() can decide whether to perform additional
+ * policy / capability checks. EACCES is used in zfs_zaccess_aces_check()
+ * to indicate access check failed due to explicit DENY entry, and so
+ * we want to avoid that here.
+ */
+static int
+zfs_zaccess_trivial(znode_t *zp, uint32_t *working_mode, cred_t *cr)
+{
+	int err, mask;
+	int unmapped = 0;
+
+	ASSERT(zp->z_pflags & ZFS_ACL_TRIVIAL);
+
+	mask = zfs_v4_to_unix(*working_mode, &unmapped);
+	if (mask == 0 || unmapped) {
+		*working_mode = unmapped;
+		return (unmapped ? SET_ERROR(EPERM) : 0);
+	}
+
+#if defined(HAVE_IOPS_PERMISSION_USERNS)
+	err = generic_permission(cr->user_ns, ZTOI(zp), mask);
+#else
+	err = generic_permission(ZTOI(zp), mask);
+#endif
+	if (err != 0) {
+		return (SET_ERROR(EPERM));
+	}
+
+	*working_mode = unmapped;
+
+	return (0);
+}
+
 static int
 zfs_zaccess_common(znode_t *zp, uint32_t v4_mode, uint32_t *working_mode,
     boolean_t *check_privs, boolean_t skipaclchk, cred_t *cr)
@@ -2450,6 +2517,9 @@ zfs_zaccess_common(znode_t *zp, uint32_t v4_mode, uint32_t *working_mode,
 		return (SET_ERROR(EPERM));
 	}
 
+	if (zp->z_pflags & ZFS_ACL_TRIVIAL)
+		return (zfs_zaccess_trivial(zp, working_mode, cr));
+
 	return (zfs_zaccess_aces_check(zp, working_mode, B_FALSE, cr));
 }