Re: kernel BUG at kernel/cred.c:434!

From: Yang Yingliang
Date: Wed Apr 17 2019 - 22:50:12 EST


Hi, Casey

On 2019/4/18 8:24, Casey Schaufler wrote:
On 4/17/2019 4:39 PM, Paul Moore wrote:
On Wed, Apr 17, 2019 at 12:27 PM Oleg Nesterov <oleg@xxxxxxxxxx> wrote:
On 04/17, Paul Moore wrote:
On Wed, Apr 17, 2019 at 10:57 AM Oleg Nesterov <oleg@xxxxxxxxxx> wrote:
On 04/17, Paul Moore wrote:
I'm tempted to simply return an error in selinux_setprocattr() if
the task's credentials are not the same as its real_cred;
What about other modules? I have no idea what smack_setprocattr() is,
but it too does prepare_creds/commit creds.

it seems that the simplest workaround should simply add the additional
cred == real_cred into proc_pid_attr_write().
Yes, that is simple, but I worry about what other LSMs might want to
do. While I believe failing if the effective creds are not the same
as the real_creds is okay for SELinux (possibly Smack too), I worry
about what other LSMs may want to do. After all,
proc_pid_attr_write() doesn't change the the creds itself, that is
something the specific LSMs do.
Yes, but if proc_pid_attr_write() is called with cred != real_cred then
something is already wrong?
True, or at least I would think so.

Looking at the current tree there are three LSMs which implement
setprocattr hooks: SELinux, Smack, and AppArmor. I know Casey has
already mentioned that he wasn't able to trigger the problem in Smack,
but looking at smack_setprocattr() I see the similar commit_creds()
usage so I would expect the same problem in Smack; what say you Casey?

I say that my test program runs without ill effect. I call acct()
with "/proc/self/attr/current", which succeeds and enables accounting
just like it is supposed to. I then have the program open
"/proc/self/attr/current" and read it, all of which goes swimmingly.
When Smack frees a cred it usually does not free any memory of its
own, so it is conceivable that I'm just getting lucky. Or, I may not
have sufficient debug enabled.

Looking at apparmor_setprocattr(), it appears that it too could end
up calling commit_creds() via aa_set_current_hat().

Since it looks like all three LSMs which implement the setprocattr
hook are vulnerable I'm open to the idea that proc_pid_attr_write() is
a better choice for the cred != read_cred check, but I would want to
make sure John and Casey are okay with that.

John?

Casey?

I'm fine with the change going into proc_pid_attr_write().
The cred != real_cred checking is not enough.

Consider this situation, when doing override, cred, real_cred and new_cred are all same:

after override_creds() cred == real_cred == new1_cred
after prepare_creds() new2_cred
after commit_creds() becasue the check is false, so cred == real_cred == new2_cred
after revert_creds() cred == new1_cred, real_cred == new2_cred

It will cause cred != real_cred finally.


Regards,
Yang