On Mon, Oct 20, 2008 at 4:04 PM, Stephen Smalley <sds@xxxxxxxxxxxxx> wrote: > On Mon, 2008-10-20 at 15:25 -0400, Daniel J Walsh wrote: >> -----BEGIN PGP SIGNED MESSAGE----- >> Hash: SHA1 >> >> Stephen Smalley wrote: >> > On Fri, 2008-10-17 at 17:18 -0400, Daniel J Walsh wrote: >> >> -----BEGIN PGP SIGNED MESSAGE----- >> >> Hash: SHA1 >> >> >> >> Stephen Smalley wrote: >> >>> On Fri, 2008-10-17 at 16:04 -0400, Eric Paris wrote: >> >>>> On Fri, 2008-10-17 at 15:39 -0400, Stephen Smalley wrote: >> >>>> >> >>>>> I don't follow the above. First, the CAP_SETUID capability controls the >> >>>>> ability to use set*uid() system calls, not to execute setuid binaries >> >>>>> (aside from a special case for shared state). Second, if some other >> >>>>> confined domain executes a setuid binary created by this user, it is >> >>>>> still limited by the permissions granted to that original confined >> >>>>> domain as far as SELinux is concerned. >> >>>> He's saying there are almost 200 domains that can run setuid apps. >> >>> As I said, CAP_SETUID isn't about whether or not you can run a setuid >> >>> app. You can do that without CAP_SETUID. Obviously since a normal >> >>> unprivileged user shell can run a setuid program in the first place. >> >>> >> >>>> Limiting the number of domains that can create new setuid apps limits >> >>>> the number of places that these domains can go. Clearly they are all >> >>>> still confined to their domain and whatever it allows, but allowing them >> >>>> to gain root priveledges may give them the ability to attack other parts >> >>> Wait - how did they gain root privileges? root privileges are >> >>> capabilities, and we control capabilities based on domain. >> >>> >> >>>> of the system normally controlled by dac. This certainly doesn't lessen >> >>>> the MAC confinement. Lets assume an audited system in which we are >> >>>> certain the only suid app untrusted users are allowed to run is ping. >> >>>> So the users have the right to run suid apps. They are protected from >> >>>> each other by DAC. >> >>>> >> >>>> Confined webadmin writes a program which is clears out other users >> >>>> public_html when they get a spurious DMCA takedown notice. He then >> >>>> (because he is a lazy bumbling idiot) makes his script suid so he does >> >>>> not have to go into his confined webadmin account constantly to delete >> >>>> users webpages. >> >>> He could also make a daemon that runs under his uid and accepts commands >> >>> via local socket. >> >>> >> >>>> Normal DAC protects user2 from being attacked by user1. Because of the >> >>>> bumbling incompetance of confined webadmin user1 can now use this setuid >> >>>> app to do things which he is allow by selinux policy but denied by >> >>>> normal DAC permissions. >> >>>> >> >>>> Why did webadmin need to make a setuid app to begin with? file caps are >> >>>> already protected by CAP_SETFCAP. Lets assume system policy says that >> >>>> su should not be setuid. Should the webadmin really be allowed to >> >>>> easily override that system policy because he wants to use su - to get >> >>>> to his confined domain instead of sudo? >> >>> He can't get to his confined domain via some other program unless policy >> >>> says he can. He can only get to some other uid that way. >> >>> >> >>>> She bumbling idiot really be >> >>>> allowed to say add o+x to su - when the system policy only really wanted >> >>>> group wheel to be able to run su? Should the bumbling idiot be able to >> >>>> remove the suid flag from a program and not be able to fix it? >> >>>> >> >>>> I think there are some real gains that can be made by limiting how >> >>>> confined admins or untrusted users can deal with suid apps. >> >>>> >> >>>> I'd probably be inclined to add changing the uid/gid/other things that >> >>>> clear setuid to be things which require this permission. I don't see a >> >>>> reason to allow my webadmin to chown su to himself... >> >>> He can't do that unless he owns su (or that itself is subject to >> >>> fowner). And chown'ing or writing to a suid app clears the suid bit >> >>> forcibly. >> >>> >> >>> Come on, guys, you can do better. >> >>> >> >> This hole discussion has actually opened me up to a new understanding. >> >> SELinux right now does not prevent the execution of setuid applications >> >> if you do not have setuid capability, it only prevents you from running >> >> apps that actually execute the setuid() call. >> >> >> >> I think this is a problem. One of the things I have been saying in my >> >> presentations is that running as staff_t will prevent you from running >> >> any setuid application unless a transition is defined. Turns out this >> >> is wrong and I was lying. >> >> >> >> # cp /usr/bin/id /usr/bin/myid >> >> # chmod 4755 /usr/bin/myid >> >> >> >> $ myid >> >> uid=3267(dwalsh) gid=3267(dwalsh) euid=0(root) groups=3267(dwalsh) >> >> context=staff_u:staff_r:staff_t:s0 >> >> >> >> So running a chmod o+s file as staff_t will run as EUID 0. I think this >> >> is something SELinux should block. At least treat this the same way a >> >> file system mounted nosetuid would. >> > >> > That would require a control on the exec path instead of chmod/setattr - >> > which is what I suggested a year ago in response to the original RFC. >> > >> Sounds like a good idea, I guess you should have pushed it. :^) >> >> > But note that gaining EUID 0 does not automatically grant privilege >> > under SELinux - your capabilities are still limited based on domain and >> > your ability to read/write even root-owned files is controlled based on >> > the (domain, type, file class) triple. >> > >> >> Understood. But we still have the risc of having a fairly wide open >> user type like staff_t tripping on a setuid app that could exculate >> privs. Or the ability for two staff_t users to attach each other >> through setuid apps. > > Ok, then I think the path forward is: > - a setsuid file permission check only applied when setting suid/sgid in > selinux_inode_setattr (and maybe when adding wider execute access, > although that's harder if you want to take ACLs into account), > > - a new process permission check in selinux_bprm_set_security() when > bprm->e_uid != current->euid or bprm->e_gid != current->egid or ! > cap_issubset(bprm->cap_post_exec_permitted, current->cap_permitted). I'll take a look. At that point is current->cap_permitted still that of the original execve process or has it already been intersected with cap_inheritable.... /me has spent all day looking at how to audit fcaps and I realized I got that wrong the first time... -Eric -- This message was distributed to subscribers of the selinux mailing list. If you no longer wish to subscribe, send mail to majordomo@xxxxxxxxxxxxx with the words "unsubscribe selinux" without quotes as the message.
