supplemental group ids (Linux)
I noticed when starting Zope as root (to get privilaged ports), but requesting suid to `nobody' (start -u nobody) the resulting processes have the correct uid and gid, but the supplemental group id list still has the appropriate value for root. This means that the Zope process could, for example, write to files that may belong to root. It's not clear whether this deserves a bug report, so I though I'd ask here instead. The fix is easy (and very lightly tested): 1) grab and install the supplemental gid package (for python) http://www.ccraig.org/software/group.c 2) patch (for 2.2.0) --- z2.py.orig Fri Jun 30 10:23:53 2000 +++ z2.py Mon Sep 4 14:33:51 2000 @@ -682,13 +682,20 @@ if type(UID) == type(""): uid = pwd.getpwnam(UID)[2] gid = pwd.getpwnam(UID)[3] + uname = UID elif type(UID) == type(1): uid = pwd.getpwuid(UID)[2] gid = pwd.getpwuid(UID)[3] + uname = pwd.getpwuid(UID)[1] else: raise KeyError try: if gid is not None: + try: + import group + group.initgroups(uname, gid) + except: + pass try: os.setgid(gid) except OSError:
Aplogies for the ignorance, but can you maybe explain the concept of supplemental group ids and give an example of how the current unpatched behavior could be subverted? On 4 Sep 2000 rugger@pangea.ca wrote:
I noticed when starting Zope as root (to get privilaged ports), but requesting suid to `nobody' (start -u nobody) the resulting processes have the correct uid and gid, but the supplemental group id list still has the appropriate value for root. This means that the Zope process could, for example, write to files that may belong to root.
It's not clear whether this deserves a bug report, so I though I'd ask here instead.
The fix is easy (and very lightly tested):
1) grab and install the supplemental gid package (for python) http://www.ccraig.org/software/group.c
2) patch (for 2.2.0)
--- z2.py.orig Fri Jun 30 10:23:53 2000 +++ z2.py Mon Sep 4 14:33:51 2000 @@ -682,13 +682,20 @@ if type(UID) == type(""): uid = pwd.getpwnam(UID)[2] gid = pwd.getpwnam(UID)[3] + uname = UID elif type(UID) == type(1): uid = pwd.getpwuid(UID)[2] gid = pwd.getpwuid(UID)[3] + uname = pwd.getpwuid(UID)[1] else: raise KeyError try: if gid is not None: + try: + import group + group.initgroups(uname, gid) + except: + pass try: os.setgid(gid) except OSError:
_______________________________________________ Zope maillist - Zope@zope.org http://lists.zope.org/mailman/listinfo/zope ** No cross posts or HTML encoding! ** (Related lists - http://lists.zope.org/mailman/listinfo/zope-announce http://lists.zope.org/mailman/listinfo/zope-dev )
Chris McDonough Digital Creations, Publishers of Zope http://www.zope.org
Chris McDonough <chrism@digicool.com> wrote:
Aplogies for the ignorance, but can you maybe explain the concept of supplemental group ids and give an example of how the current unpatched behavior could be subverted?
I can try... Supplemental gids are useful for allowing a user to belong to more than one group, or maybe to more than one project in normal parlance. This is normally effected by listing the uid opposite more than one group in /etc/group. The login process issues the initgroups(3) call to install these supplemental groups, which are inherited by all processes forked from the login shell. The problem is comes when you change user ids; for example what I saw with Zope (start -u nobody) was: before change after change ============= ============ user id root nobody group id root nobody sup id(s) root root Thus the process has group access privilages for nobody (correct) and root (bad) in unpatched Zope. I cannot give you an exploit based on this -- my knowledge of Zope is not deep enough -- and in a bug free world there probably would be no exploit. But the reason for running as nobody, I think, is to contain damage should an exploit be found. For that reason, it would seem reasonable to change the supplemental gids too. I saw this on Linux; supplemental groups come from the BSD tradition, so you likely will find the same situation on *BSD, Solaris, etc.
On 4 Sep 2000 rugger@pangea.ca wrote: I see... well, maybe we can take a look at it. In the meantime, if you figure out a patch that doesn't rely on an external program, let me know... Thanks, C
Chris McDonough <chrism@digicool.com> wrote:
Aplogies for the ignorance, but can you maybe explain the concept of supplemental group ids and give an example of how the current unpatched behavior could be subverted?
I can try...
Supplemental gids are useful for allowing a user to belong to more than one group, or maybe to more than one project in normal parlance. This is normally effected by listing the uid opposite more than one group in /etc/group. The login process issues the initgroups(3) call to install these supplemental groups, which are inherited by all processes forked from the login shell.
The problem is comes when you change user ids; for example what I saw with Zope (start -u nobody) was:
before change after change ============= ============ user id root nobody group id root nobody sup id(s) root root
Thus the process has group access privilages for nobody (correct) and root (bad) in unpatched Zope.
I cannot give you an exploit based on this -- my knowledge of Zope is not deep enough -- and in a bug free world there probably would be no exploit. But the reason for running as nobody, I think, is to contain damage should an exploit be found. For that reason, it would seem reasonable to change the supplemental gids too.
I saw this on Linux; supplemental groups come from the BSD tradition, so you likely will find the same situation on *BSD, Solaris, etc.
_______________________________________________ Zope maillist - Zope@zope.org http://lists.zope.org/mailman/listinfo/zope ** No cross posts or HTML encoding! ** (Related lists - http://lists.zope.org/mailman/listinfo/zope-announce http://lists.zope.org/mailman/listinfo/zope-dev )
Chris McDonough Digital Creations, Publishers of Zope http://www.zope.org
| > I saw this on Linux; supplemental groups come from the BSD tradition, | > so you likely will find the same situation on *BSD, Solaris, etc. Sorry I missed the start of the thread, but, I can weigh in on this point. Using -u user under FreeBSD gives you the Primary Group for the user you have requested. Supplemental groups are available only for that user, not the user you ran Zope as. -- Totally Holistic Enterprises Internet| P:+61 7 3870 0066 | Andrew Milton The Internet (Aust) Pty Ltd | F:+61 7 3870 4477 | ACN: 082 081 472 ABN: 83 082 081 472 | M:+61 416 022 411 | Carpe Daemon PO Box 837 Indooroopilly QLD 4068 |akm@theinternet.com.au|
Andrew Kenneth Milton <akm@mail.theinternet.com.au> wrote:
| > I saw this on Linux; supplemental groups come from the BSD tradition, | > so you likely will find the same situation on *BSD, Solaris, etc.
Sorry I missed the start of the thread, but, I can weigh in on this point.
Using -u user under FreeBSD gives you the Primary Group for the user you have requested. Supplemental groups are available only for that user, not the user you ran Zope as.
Interesting. NetBSD has the Linux behaviour. The Apache CGI wrapper does setgid and initgroups, in that order.
Kip Rugger wrote:
Chris McDonough <chrism@digicool.com> wrote:
Aplogies for the ignorance, but can you maybe explain the concept of supplemental group ids and give an example of how the current unpatched behavior could be subverted?
I can try...
Supplemental gids are useful for allowing a user to belong to more than one group, or maybe to more than one project in normal parlance. This is normally effected by listing the uid opposite more than one group in /etc/group. The login process issues the initgroups(3) call to install these supplemental groups, which are inherited by all processes forked from the login shell. The problem is comes when you change user ids; for example what I saw with Zope (start -u nobody) was:
before change after change ============= ============ user id root nobody group id root nobody sup id(s) root root
Would you mind describing how you determine this? -- Do not meddle in the affairs of sysadmins, for they are easy to annoy, and have the root password.
Bill Anderson <bill@libc.org> wrote:
Kip Rugger wrote:
Chris McDonough <chrism@digicool.com> wrote:
Aplogies for the ignorance, but can you maybe explain the concept of supplemental group ids and give an example of how the current unpatched behavior could be subverted?
I can try...
Supplemental gids are useful for allowing a user to belong to more than one group, or maybe to more than one project in normal parlance. This is normally effected by listing the uid opposite more than one group in /etc/group. The login process issues the initgroups(3) call to install these supplemental groups, which are inherited by all processes forked from the login shell. The problem is comes when you change user ids; for example what I saw with Zope (start -u nobody) was:
before change after change ============= ============ user id root nobody group id root nobody sup id(s) root root
Would you mind describing how you determine this?
[/proc] $ cat /proc/90/status Name: junkbuster State: S (sleeping) Pid: 90 PPid: 1 Uid: 101 101 101 101 Gid: 101 101 101 101 Groups: 101 <------------------------------ supplemental groups VmSize: 1348 kB VmLck: 0 kB VmRSS: 436 kB VmData: 192 kB VmStk: 84 kB VmExe: 92 kB VmLib: 952 kB SigPnd: 0000000000000000 SigBlk: 0000000000000000 SigIgn: 0000000080011006 SigCgt: 0000000000000000 CapInh: 00000000fffffeff CapPrm: 0000000000000000 CapEff: 0000000000000000 On my machine 101 is uid and gid for nobody; as you can see junkbuster is correctly sandboxed. For unmodified Zope, you'll see a zero in the indicated line (or possibly several values if root belongs to several groups like `wheel' on your system).
+-------[ Kip Rugger ]---------------------- | | On my machine 101 is uid and gid for nobody; as you can see | junkbuster is correctly sandboxed. For unmodified Zope, you'll | see a zero in the indicated line (or possibly several values | if root belongs to several groups like `wheel' on your system). Have you actually tried to change to that group? -- Totally Holistic Enterprises Internet| P:+61 7 3870 0066 | Andrew Milton The Internet (Aust) Pty Ltd | F:+61 7 3870 4477 | ACN: 082 081 472 ABN: 83 082 081 472 | M:+61 416 022 411 | Carpe Daemon PO Box 837 Indooroopilly QLD 4068 |akm@theinternet.com.au|
Kip Rugger wrote:
Bill Anderson <bill@libc.org> wrote:
Kip Rugger wrote:
Chris McDonough <chrism@digicool.com> wrote:
Aplogies for the ignorance, but can you maybe explain the concept of supplemental group ids and give an example of how the current unpatched behavior could be subverted?
I can try...
Supplemental gids are useful for allowing a user to belong to more than one group, or maybe to more than one project in normal parlance. This is normally effected by listing the uid opposite more than one group in /etc/group. The login process issues the initgroups(3) call to install these supplemental groups, which are inherited by all processes forked from the login shell. The problem is comes when you change user ids; for example what I saw with Zope (start -u nobody) was:
before change after change ============= ============ user id root nobody group id root nobody sup id(s) root root
Would you mind describing how you determine this?
[/proc] $ cat /proc/90/status Name: junkbuster State: S (sleeping) Pid: 90 PPid: 1 Uid: 101 101 101 101 Gid: 101 101 101 101 Groups: 101 <------------------------------ supplemental groups
....
On my machine 101 is uid and gid for nobody; as you can see junkbuster is correctly sandboxed. For unmodified Zope, you'll see a zero in the indicated line (or possibly several values if root belongs to several groups like `wheel' on your system).
OK, something is not quite right here. On my unmodified zope, it is properly 'sandboxed'. Perhaps it is the use of the explicit '-u nobody'? I don't do that on my system, which causes Zope to run as nobody implicitly. (When started as root, unless told otherwise, zope will switch to nobody). Try running without the 'u nobody switch, and see what happens. Just out of curiousity. -- Do not meddle in the affairs of sysadmins, for they are easy to annoy, and have the root password.
OK, something is not quite right here. On my unmodified zope, it is properly 'sandboxed'. Perhaps it is the use of the explicit '-u nobody'? I don't do that on my system, which causes Zope to run as nobody implicitly.
(When started as root, unless told otherwise, zope will switch to nobody).
Try running without the 'u nobody switch, and see what happens. Just out of curiousity.
No difference. I think the point is that Zope does not make any initgroups(3) calls; this will be a problem if the particular system needs it. I have two such systems: Linux 2.2.16 + glibc-2.1.2 NetBSD 1.4 Having reviewed the kernel and libc sources in both cases, I am convinced that set*gid and {init,set,get}groups operate totally independently. At minimum, initgroups is used by login/su to set the primary gid found in /etc/passwd, plus any additional gids associated with the uid in /etc/group, as supplemental gids. Thus, even if there are no supplemental gids in /etc/group, you still have the primary gid in the kernel's list of supplementals. So the primary gid occurs initially in 3 places: the real and effective gids, and one of the supplemental gids. You must get all 3; setgid for real and eff, initgroups for sup. (Additionally in linux you have the `saved' gid and the fsgid, but setgid will modify them.) Under this hypothesis, my question is how could _your_ system work? Why is it that you don't have the original primary gid lingering in the supplemental list?
Kip Rugger wrote:
OK, something is not quite right here. On my unmodified zope, it is properly 'sandboxed'. Perhaps it is the use of the explicit '-u nobody'? I don't do that on my system, which causes Zope to run as nobody implicitly.
(When started as root, unless told otherwise, zope will switch to nobody).
Try running without the 'u nobody switch, and see what happens. Just out of curiousity.
No difference.
I think the point is that Zope does not make any initgroups(3) calls; this will be a problem if the particular system needs it.
I have two such systems:
Linux 2.2.16 + glibc-2.1.2 NetBSD 1.4
...
Under this hypothesis, my question is how could _your_ system work? Why is it that you don't have the original primary gid lingering in the supplemental list?
Not sure. Here is my setup: glibc 2.1.3 Kernel 2.2.15 heavilly modified Redhat 6.2 base. Perhaps it is the kernel? I also have a 2.2.16 (2.1.3 glibc) kernelled machine which exhibits the behavior you see on yours.. I can try it on a 2.2.4test6 kernel too ... -- Do not meddle in the affairs of sysadmins, for they are easy to annoy, and have the root password.
After some digging, it appears that this is a really good find. Thanks very much for reporting it. I am going to add a collector item with your message verbatim. Thanks very much! - C On 4 Sep 2000 rugger@pangea.ca wrote:
I noticed when starting Zope as root (to get privilaged ports), but requesting suid to `nobody' (start -u nobody) the resulting processes have the correct uid and gid, but the supplemental group id list still has the appropriate value for root. This means that the Zope process could, for example, write to files that may belong to root.
It's not clear whether this deserves a bug report, so I though I'd ask here instead.
The fix is easy (and very lightly tested):
1) grab and install the supplemental gid package (for python) http://www.ccraig.org/software/group.c
2) patch (for 2.2.0)
--- z2.py.orig Fri Jun 30 10:23:53 2000 +++ z2.py Mon Sep 4 14:33:51 2000 @@ -682,13 +682,20 @@ if type(UID) == type(""): uid = pwd.getpwnam(UID)[2] gid = pwd.getpwnam(UID)[3] + uname = UID elif type(UID) == type(1): uid = pwd.getpwuid(UID)[2] gid = pwd.getpwuid(UID)[3] + uname = pwd.getpwuid(UID)[1] else: raise KeyError try: if gid is not None: + try: + import group + group.initgroups(uname, gid) + except: + pass try: os.setgid(gid) except OSError:
_______________________________________________ Zope maillist - Zope@zope.org http://lists.zope.org/mailman/listinfo/zope ** No cross posts or HTML encoding! ** (Related lists - http://lists.zope.org/mailman/listinfo/zope-announce http://lists.zope.org/mailman/listinfo/zope-dev )
Chris McDonough Digital Creations, Publishers of Zope http://www.zope.org
participants (5)
-
Andrew Kenneth Milton -
Bill Anderson -
Chris McDonough -
kbr@pangea.ca -
rugger@pangea.ca