Full relabel audit event

The attached patch implements the full relabel audit event (Ie. an
audit event occurs when a full relabel occurs, ie. when /.autorelabel exists at boot).
Note that although the code is correct, this patch doesn't actually
work due to kernel bugs[1].

It'll be in Fedora development as part of policycoreutils-1.30.10-3
onwards.

[1] see the thread on linux-audit if you want the details.

James Antill <[email protected]>

diff -ru policycoreutils-1.30.10-orig/setfiles/setfiles.c policycoreutils-1.30.10/setfiles/setfiles.c
@@ -75,6 +75,11 @@
 #include <selinux/selinux.h>
 #include <syslog.h> 
 #include <libgen.h>
+#include <libaudit.h>
+
+#ifndef AUDIT_FS_RELABEL
+#define AUDIT_FS_RELABEL 2309
+#endif
 
 static int add_assoc = 1;
 static FILE *outfile=NULL;

@@ -395,7 +400,7 @@
 
 	/* trim trailing /, if present */
 	len = strlen(rootpath);
-	while ('/' == rootpath[len - 1])
+	while (len && ('/' == rootpath[len - 1]))
 		rootpath[--len] = 0;
 	rootpathlen = len;
 }

@@ -443,11 +448,35 @@
   return 0;
 }
 
+static void maybe_audit_mass_relabel(int done_root, int errs)
+{
+  int audit_fd = -1;
+  int rc = 0;
+  
+  if (!done_root) /* only audit a forced full relabel */
+    return;
+  
+  audit_fd = audit_open();
+
+  if (audit_fd < 0) {
+    fprintf(stderr, "Error connecting to audit system.\n");
+    return;
+  }
+
+  rc = audit_log_user_message(audit_fd, AUDIT_FS_RELABEL,
+                              "op=mass relabel", NULL, NULL, NULL, !errs);
+  if (rc <= 0) {
+    fprintf(stderr, "Error sending audit message: %s.\n", strerror(errno));
+  }
+  audit_close(audit_fd);
+}
+
 int main(int argc, char **argv)
 {
 	struct stat sb;
 	int opt, rc, i;
-
+        int done_root = 0; /* have we processed the / directory as an arg */
+        
 	memset(excludeArray,0, sizeof(excludeArray));
 
 	/* Validate all file contexts during matchpathcon_init. */

@@ -618,6 +647,8 @@
 	}
 	else for (; optind < argc; optind++)
 	{
+                done_root |= !strcmp(argv[optind], "/");
+                
 		if (NULL != rootpath) {
 			qprintf("%s:  labeling files, pretending %s is /\n",
 				argv[0], rootpath);

@@ -648,6 +679,7 @@
 				fprintf(stderr,
 				"%s:  error while labeling files under %s\n",
 				argv[0], argv[optind]);
+				maybe_audit_mass_relabel(done_root, 1);
 				exit(1);
 			}
 		}

@@ -664,6 +696,8 @@
 		matchpathcon_filespec_destroy();
 	}
 
+        maybe_audit_mass_relabel(done_root, 0);
+
 	if (warn_no_match)
 		matchpathcon_checkmatches(argv[0]);
 

@@ -7,6 +7,7 @@
 CFLAGS = -Werror -Wall -W 
 override CFLAGS += -D_FILE_OFFSET_BITS=64 -I$(PREFIX)/include
 LDLIBS = -lselinux -lsepol -L$(LIBDIR)
+LDLIBS += -laudit
 
 all: setfiles
 

--
redhat-lspp mailing list
[email protected]
https://www.redhat.com/mailman/listinfo/redhat-lspp

Hmmm...what is it that you actually want to do here? If you only care about auditing autorelabel events, then I'd suggest generating the audit message from the autorelabel portion of rc.sysinit (via a helper, I suppose), not from setfiles itself. If you want to audit all full relabels, then you need to instrument more than setfiles (e.g. restorecon -R / works just as well), and of course, you potentially need to do something at the kernel level with audit filters or auditallow rules in policy if you truly want to capture all relabels. And, of course, just auditing it when they happen to pass "/" as an argument isn't very reliable.

Not sure which thread you are referring to; I don't see prior discussion of a relabel audit event in the linux-audit archives.

Stephen Smalley
National Security Agency

This is all that we care about, but the solution of creating a helper
to just be called before setfiles was considered suboptimal against just putting the code inside setfiles (I know Steve is very much against anything which acts like logger for the audit subsystem).

The thread is for the kernel problem that makes the above patch not
actually work, see the thread "Re: audit 1.2.2 released".

James Antill <[email protected]>

I don't see the point when a) you only want it in that one case, b) it doesn't prevent trivial bypass in any way (e.g. by using restorecon, by rolling your own program to do it, by running setfiles on /* rather than just /, ...), and c) you aren't capturing any information that cannot be determined by the caller of setfiles in the first place (just the fact of a mass relabel and the final exit status).

Note btw that setfiles already provides three different ways to log actual changes in file contexts, the original -v verbose mode, and the -l (log via syslog) and -o <file> (log to file) modes introduced later by Red Hat. That at least provides detailed information that the caller couldn't determine otherwise.

Stephen Smalley
National Security Agency

We do this already in several places. For example, we instrumented usermod, but not chage. It was documented in the Security Target that usermod should be used to alter user account attributes.

Its not meant to be bulletproof. Its purpose is to document that a full relabel has occurred before any user can log in. During the boot process, no one can log in so nothing evil should happen.

It was determined that we only need 1 record, not all the changes.

-Steve

--
redhat-lspp mailing list
[email protected]
https://www.redhat.com/mailman/listinfo/redhat-lspp

We need to meet the requirements for LSPP where there is a relabel on boot, but we do not want a record for each file that was touched. It was discussed on the LSPP telecon a while back that just one record was sufficient.

This is a shell script and cannot connect to libaudit.

We get relabels by monitoring the setxattr syscall. But during bootup before going interactive, we just want 1 message.

-Steve

--
redhat-lspp mailing list
[email protected]
https://www.redhat.com/mailman/listinfo/redhat-lspp