On 1/24/2020 12:16 PM, Stephen Smalley wrote: > On 1/24/20 2:28 PM, Casey Schaufler wrote: >> On 1/24/2020 8:20 AM, Stephen Smalley wrote: >>> On 1/24/20 9:42 AM, Stephen Smalley wrote: >>>> On 1/23/20 7:23 PM, Casey Schaufler wrote: >>>>> Add an entry /proc/.../attr/context which displays the full >>>>> process security "context" in compound format:' >>>>> lsm1\0value\0lsm2\0value\0... >>>>> This entry is not writable. >>>>> >>>>> Signed-off-by: Casey Schaufler <casey@xxxxxxxxxxxxxxxx> >>>>> Cc: linux-api@xxxxxxxxxxxxxxx >>>> >>>> As previously discussed, there are issues with AppArmor's implementation of getprocattr() particularly around the trailing newline that dbus-daemon and perhaps others would like to see go away in any new interface. Hence, I don't think we should implement this new API using the existing getprocattr() hook lest it also be locked into the current behavior forever. >>> >>> Also, it would be good if whatever hook is introduced to support /proc/pid/attr/context could also be leveraged by the SO_PEERCONTEXT implementation in the future so that we are guaranteed a consistent result between the two interfaces, unlike the current situation for /proc/self/attr/current versus SO_PEERSEC. >> >> I don't believe that a new hook is necessary, and that introducing one >> just to deal with a '\n' would be pedantic. We really have two rational >> options. AppArmor could drop the '\n' from their "context". Or, we can >> simply document that the /proc/pid/attr/context interface will trim any >> trailing whitespace. I understand that this would be a break from the >> notion that the LSM infrastructure does not constrain what a module uses >> for its own data. On the other hand, we have been saying that "context"s >> are strings, and ignoring trailing whitespace is usual behavior for >> strings. > > Well, you can either introduce a new common underlying hook for use by /proc/pid/attr/context and SO_PEERCONTEXT to produce the string that is to be returned to userspace (in order to guarantee consistency in format and allowing them to be directly compared, which I think is what the dbus maintainers wanted), The getprocattr() hooks are already required to provide a nul terminated string. The behavior of /proc/self/attr/current with regard to trailing whitespace or the terminating nul is left to the security module. The behavior of /proc/self/attr/context is new, and as such it can be defined to be reasonable and consistent. > or you can modify every security module to provide that guarantee in its existing getprocattr and getpeersec* hook functions (SELinux already provides this guarantee; Smack and AppArmor produce slightly different results with respect to \0 and/or \n), or you can have the framework trim the values it gets from the security modules before composing them. Changing "every security module" turns out to be trivial, really. The security_getprocattr() interface doesn't change, the getprocattr hooks get a bool argument indicating if newlines are allowed, the AppArmor code changes to check the bool and security_getprocattr() sets to bool specially for the "context" case. The getpeersec() interface use for SO_PEERCONTEXT isn't any more difficult. The security modules are free to include a terminating nul or not in the existing use case, but a bool can tell them what the behavior must be when we get to that interface. It's not like there are all that many security modules to bring in line at this point. Establishing sane rules of behavior is overdue. With the current set of new modules sniffing at the door it really is time to tighten these up. > But you need to do one of those things before this interface gets merged upstream. > > Aside from the trailing newline and \0 issues, AppArmor also has a whitespace-separated (mode) field that may or may not be present in the contexts it presently returns, ala "/usr/sbin/cupsd (enforce)". Not sure what they want for the new interfaces. Using '\0' as a field terminator addresses that issue. That's a primary motivator.
