SELinux: work around filesystems which call d_instantiate before setting inode mode

One filesystem in particular, CIFS, is known to call d_instantiate before setting the mode for for some operations. It will create a dentry for any children when doing a directory search and thereby is
calling d_instantiate. But it does so before setting the inode mode
for the child inodes. Thus the selinux sclass in the corresponding inode security struct is set incorrectly to always believe these inodes are regular files. Then when operations are performed on these inodes at a later point in time SELinux will deny operations which may be allowed for the correct class but not for files or SELinux may check for access permissions to do operations which do not even pertain to the 'file' class. An example would be the user may attempt to remove a subdirectory which would need SELinux permissions to rmdir what it believes is a 'file.' rmdir'ing a regular file doesn't make sense and is obviously not properly defined. What this patch does is to recalculate the sclass for each inode on each permission check. Thus if a filesystem decided to later (after the d_instantiate) set the mode bits (as CIFS does) we will make the correct security checks. We also output a warning message letting the user know that they have a filesystem which doing operations in a questionable order. (Questionable because after calling d_instantiate the new inode may be available to other threads through the dentry cache with the mode set improperly)

I believe that the CIFS people were contacted twice trying to get this changed but we want SELinux users to be able to work now and we want to know if any other filesystem uses this same ordering or choose to implement it in the future.

Signed-off-by Eric Paris <[email protected]>

 security/selinux/hooks.c       |   27 +++++++++++++++++++++++++++
 security/selinux/include/avc.h |    2 ++
 security/selinux/avc.c         |   14 +++++++++++++-
 3 files changed, 42 insertions(+), 1 deletions(-)

security/selinux/hooks.c +27 -0

index 5d1b8c7..5527aec 100644
@@ -1066,6 +1066,25 @@ static int task_has_system(struct task_s
 			    SECCLASS_SYSTEM, perms, NULL);
 }
 
+/* 
+ * Update the sclass of an inode.  This shouldn't ever do anything unless a FS
+ * actually called d_instantiate before it set the i_mode.
+ */
+static inline void inode_update_sclass(struct inode *inode)
+{
+	struct inode_security_struct *isec = inode->i_security;
+	if (isec->sclass == SECCLASS_FILE)
+	{
+		isec->sclass = inode_mode_to_security_class(inode->i_mode);
+		if (unlikely(isec->sclass != SECCLASS_FILE) && printk_ratelimit())
+			printk(KERN_WARNING "SELinux: Inode on a %s filesystem "
+				"with sclass=file but should have been sclass="
+				"%s, fixing up this issue\n", 
+				inode->i_sb->s_type->name, 
+				avc_class_to_string(isec->sclass));
+	}
+}
+
 /* Check whether a task has a particular permission to an inode.
    The 'adp' parameter is optional and allows other audit
    data to be passed (e.g. the dentry). */

@@ -1081,6 +1100,8 @@ static int inode_has_perm(struct task_st
 	tsec = tsk->security;
 	isec = inode->i_security;
 
+	inode_update_sclass(inode);
+
 	if (!adp) {
 		adp = &ad;
 		AVC_AUDIT_DATA_INIT(&ad, FS);

@@ -1220,6 +1241,8 @@ static int may_link(struct inode *dir,
 	dsec = dir->i_security;
 	isec = dentry->d_inode->i_security;
 
+	inode_update_sclass(dentry->d_inode);
+
 	AVC_AUDIT_DATA_INIT(&ad, FS);
 	ad.u.fs.dentry = dentry;
 

@@ -1266,6 +1289,8 @@ static inline int may_rename(struct inod
 	old_is_dir = S_ISDIR(old_dentry->d_inode->i_mode);
 	new_dsec = new_dir->i_security;
 
+	inode_update_sclass(old_dir);
+
 	AVC_AUDIT_DATA_INIT(&ad, FS);
 
 	ad.u.fs.dentry = old_dentry;

@@ -2260,6 +2285,8 @@ static int selinux_inode_setxattr(struct
 	if ((current->fsuid != inode->i_uid) && !capable(CAP_FOWNER))
 		return -EPERM;
 
+	inode_update_sclass(inode);
+
 	AVC_AUDIT_DATA_INIT(&ad,FS);
 	ad.u.fs.dentry = dentry;
 

security/selinux/include/avc.h +2 -0

index 960ef18..043d479 100644
@@ -125,6 +125,8 @@ int avc_add_callback(int (*callback)(u32
 		     u32 events, u32 ssid, u32 tsid,
 		     u16 tclass, u32 perms);
 
+const char *avc_class_to_string(u16 tclass);
+
 /* Exported to selinuxfs */
 int avc_get_hash_stats(char *page);
 extern unsigned int avc_cache_threshold;

security/selinux/avc.c +13 -1

index a300702..88bba69 100644
@@ -218,7 +218,7 @@ static void avc_dump_query(struct audit_
 		audit_log_format(ab, " tcontext=%s", scontext);
 		kfree(scontext);
 	}
-	audit_log_format(ab, " tclass=%s", class_to_string[tclass]);
+	audit_log_format(ab, " tclass=%s", avc_class_to_string(tclass));
 }
 
 /**

@@ -913,3 +913,15 @@ int avc_has_perm(u32 ssid, u32 tsid, u16
 	avc_audit(ssid, tsid, tclass, requested, &avd, rc, auditdata);
 	return rc;
 }
+
+/**
+ * avc_class_to_string - return a human readable string given an object class.
+ * @tclass: the target class we wish to translate
+ *
+ * Simply take the target object class passed to us and return the human
+ * readable string associated with that class
+ */
+const char *avc_class_to_string(u16 tclass)
+{
+	return class_to_string[tclass];
+}

Quoted 1 line

Signed-off-by Eric Paris <[email protected]>

Acked-by Stephen Smalley <[email protected]>